update version to 3.3.0post1

Currently translated at 98.0% (1186 of 1210 strings)

translationBot(ui): update translation (Chinese (Simplified))

Currently translated at 98.0% (1179 of 1203 strings)

translationBot(ui): update translation (Chinese (Simplified))

Currently translated at 97.9% (1175 of 1199 strings)

Co-authored-by: Surisen <zhonghx0804@outlook.com>
Translate-URL: https://hosted.weblate.org/projects/invokeai/web-ui/zh_Hans/
Translation: InvokeAI/Web UI

docs/contributing/LOCAL_DEVELOPMENT.md

@@ -47,34 +47,9 @@ pip install ".[dev,test]"
 These are optional groups of packages which are defined within the `pyproject.toml`
 and will be required for testing the changes you make to the code.
 
-### Running Tests
-
-We use [pytest](https://docs.pytest.org/en/7.2.x/) for our test suite. Tests can
-be found under the `./tests` folder and can be run with a single `pytest`
-command. Optionally, to review test coverage you can append `--cov`.
-
-```zsh
-pytest --cov
-```
-
-Test outcomes and coverage will be reported in the terminal. In addition a more
-detailed report is created in both XML and HTML format in the `./coverage`
-folder. The HTML one in particular can help identify missing statements
-requiring tests to ensure coverage. This can be run by opening
-`./coverage/html/index.html`.
-
-For example.
-
-```zsh
-pytest --cov; open ./coverage/html/index.html
-```
-
-??? info "HTML coverage report output"
-
-    ![html-overview](../assets/contributing/html-overview.png)
-
-    ![html-detail](../assets/contributing/html-detail.png)
+### Tests
 
+See the [tests documentation](./TESTS.md) for information about running and writing tests.
 ### Reloading Changes
 
 Experimenting with changes to the Python source code is a drag if you have to re-start the server —
@@ -167,6 +142,23 @@ and so you'll have access to the same python environment as the InvokeAI app.
 
 This is _super_ handy.
 
+#### Enabling Type-Checking with Pylance
+
+We use python's typing system in InvokeAI. PR reviews will include checking that types are present and correct. We don't enforce types with `mypy` at this time, but that is on the horizon.
+
+Using a code analysis tool to automatically type check your code (and types) is very important when writing with types. These tools provide immediate feedback in your editor when types are incorrect, and following their suggestions lead to fewer runtime bugs.
+
+Pylance, installed at the beginning of this guide, is the de-facto python LSP (language server protocol). It provides type checking in the editor (among many other features). Once installed, you do need to enable type checking manually:
+
+- Open a python file
+- Look along the status bar in VSCode for `{ } Python`
+- Click the `{ }`
+- Turn type checking on - basic is fine
+
+You'll now see red squiggly lines where type issues are detected. Hover your cursor over the indicated symbols to see what's wrong.
+
+In 99% of cases when the type checker says there is a problem, there really is a problem, and you should take some time to understand and resolve what it is pointing out.
+
 #### Debugging configs with `launch.json`
 
 Debugging configs are managed in a `launch.json` file. Like most VSCode configs,

docs/contributing/TESTS.md

@@ -0,0 +1,89 @@
+# InvokeAI Backend Tests
+
+We use `pytest` to run the backend python tests. (See [pyproject.toml](/pyproject.toml) for the default `pytest` options.)
+
+## Fast vs. Slow
+All tests are categorized as either 'fast' (no test annotation) or 'slow' (annotated with the `@pytest.mark.slow` decorator).
+
+'Fast' tests are run to validate every PR, and are fast enough that they can be run routinely during development.
+
+'Slow' tests are currently only run manually on an ad-hoc basis. In the future, they may be automated to run nightly. Most developers are only expected to run the 'slow' tests that directly relate to the feature(s) that they are working on.
+
+As a rule of thumb, tests should be marked as 'slow' if there is a chance that they take >1s (e.g. on a CPU-only machine with slow internet connection). Common examples of slow tests are tests that depend on downloading a model, or running model inference.
+
+## Running Tests
+
+Below are some common test commands:
+```bash
+# Run the fast tests. (This implicitly uses the configured default option: `-m "not slow"`.)
+pytest tests/
+
+# Equivalent command to run the fast tests.
+pytest tests/ -m "not slow"
+
+# Run the slow tests.
+pytest tests/ -m "slow"
+
+# Run the slow tests from a specific file.
+pytest tests/path/to/slow_test.py -m "slow"
+
+# Run all tests (fast and slow).
+pytest tests -m ""
+```
+
+## Test Organization
+
+All backend tests are in the [`tests/`](/tests/) directory. This directory mirrors the organization of the `invokeai/` directory. For example, tests for `invokeai/model_management/model_manager.py` would be found in `tests/model_management/test_model_manager.py`.
+
+TODO: The above statement is aspirational. A re-organization of legacy tests is required to make it true.
+
+## Tests that depend on models
+
+There are a few things to keep in mind when adding tests that depend on models.
+
+1. If a required model is not already present, it should automatically be downloaded as part of the test setup.
+2. If a model is already downloaded, it should not be re-downloaded unnecessarily.
+3. Take reasonable care to keep the total number of models required for the tests low. Whenever possible, re-use models that are already required for other tests. If you are adding a new model, consider including a comment to explain why it is required/unique.
+
+There are several utilities to help with model setup for tests. Here is a sample test that depends on a model:
+```python
+import pytest
+import torch
+
+from invokeai.backend.model_management.models.base import BaseModelType, ModelType
+from invokeai.backend.util.test_utils import install_and_load_model
+
+@pytest.mark.slow
+def test_model(model_installer, torch_device):
+    model_info = install_and_load_model(
+        model_installer=model_installer,
+        model_path_id_or_url="HF/dummy_model_id",
+        model_name="dummy_model",
+        base_model=BaseModelType.StableDiffusion1,
+        model_type=ModelType.Dummy,
+    )
+
+    dummy_input = build_dummy_input(torch_device)
+
+    with torch.no_grad(), model_info as model:
+        model.to(torch_device, dtype=torch.float32)
+        output = model(dummy_input)
+
+    # Validate output...
+
+```
+
+## Test Coverage
+
+To review test coverage, append `--cov` to your pytest command:
+```bash
+pytest tests/ --cov
+```
+
+Test outcomes and coverage will be reported in the terminal. In addition, a more detailed report is created in both XML and HTML format in the `./coverage` folder. The HTML output is particularly helpful in identifying untested statements where coverage should be improved. The HTML report can be viewed by opening `./coverage/html/index.html`.
+
+??? info "HTML coverage report output"
+
+    ![html-overview](../assets/contributing/html-overview.png)
+
+    ![html-detail](../assets/contributing/html-detail.png)

docs/contributing/contribution_guides/development.md

@@ -12,7 +12,7 @@ To get started, take a look at our [new contributors checklist](newContributorCh
 Once you're setup, for more information, you can review the documentation specific to your area of interest:
 
 * #### [InvokeAI Architecure](../ARCHITECTURE.md)
-* #### [Frontend Documentation](development_guides/contributingToFrontend.md)
+* #### [Frontend Documentation](./contributingToFrontend.md)
 * #### [Node Documentation](../INVOCATIONS.md)
 * #### [Local Development](../LOCAL_DEVELOPMENT.md)
 
@@ -38,9 +38,9 @@ There are two paths to making a development contribution:
 
 If you need help, you can ask questions in the [#dev-chat](https://discord.com/channels/1020123559063990373/1049495067846524939) channel of the Discord.
 
-For frontend related work, **@pyschedelicious** is the best person to reach out to. 
+For frontend related work, **@psychedelicious** is the best person to reach out to. 
 
-For backend related work, please reach out to **@blessedcoolant**, **@lstein**, **@StAlKeR7779** or **@pyschedelicious**.
+For backend related work, please reach out to **@blessedcoolant**, **@lstein**, **@StAlKeR7779** or **@psychedelicious**.
 
 
 ## **What does the Code of Conduct mean for me?**

docs/contributing/contribution_guides/documentation.md

@@ -10,4 +10,4 @@ When updating or creating documentation, please keep in mind InvokeAI is a tool
 
 ## Help & Questions
 
-Please ping @imic1 or @hipsterusername in the [Discord](https://discord.com/channels/1020123559063990373/1049495067846524939) if you have any questions.
+Please ping @imic or @hipsterusername in the [Discord](https://discord.com/channels/1020123559063990373/1049495067846524939) if you have any questions.

docs/features/CONFIGURATION.md

@@ -159,7 +159,7 @@ groups in `invokeia.yaml`:
 | `host`     | `localhost`      | Name or IP address of the network interface that the web server will listen on  |
 | `port`     | `9090`           | Network port number that the web server will listen on  |
 | `allow_origins`  | `[]`       | A list of host names or IP addresses that are allowed to connect to the InvokeAI API in the format `['host1','host2',...]` |
-| `allow_credentials | `true`   | Require credentials for a foreign host to access the InvokeAI API (don't change this) |
+| `allow_credentials` | `true`   | Require credentials for a foreign host to access the InvokeAI API (don't change this) |
 | `allow_methods` | `*`         | List of HTTP methods ("GET", "POST") that the web server is allowed to use when accessing the API  |
 | `allow_headers` | `*`         | List of HTTP headers that the web server will accept when accessing the API  |
 

docs/features/CONTROLNET.md

@@ -1,13 +1,11 @@
 ---
-title: ControlNet
+title: Control Adapters
 ---
 
-# :material-loupe: ControlNet
+# :material-loupe: Control Adapters
 
 ## ControlNet
 
-ControlNet
-
 ControlNet is a powerful set of features developed by the open-source
 community (notably, Stanford researcher
 [**@ilyasviel**](https://github.com/lllyasviel)) that allows you to
@@ -20,7 +18,7 @@ towards generating images that better fit your desired style or
 outcome.
 
 
-### How it works
+#### How it works
 
 ControlNet works by analyzing an input image, pre-processing that
 image to identify relevant information that can be interpreted by each
@@ -30,7 +28,7 @@ composition, or other aspects of the image to better achieve a
 specific result.
 
 
-### Models
+#### Models
 
 InvokeAI provides access to a series of ControlNet models that provide
 different effects or styles in your generated images.  Currently
@@ -96,6 +94,8 @@ A model that generates normal maps from input images, allowing for more realisti
 **Image Segmentation**: 
 A model that divides input images into segments or regions, each of which corresponds to a different object or part of the image. (More details coming soon)
 
+**QR Code Monster**:
+A model that helps generate creative QR codes that still scan. Can also be used to create images with text, logos or shapes within them. 
 
 **Openpose**: 
 The OpenPose control model allows for the identification of the general pose of a character by pre-processing an existing image with a clear human structure. With advanced options, Openpose can also detect the face or hands in the image. 
@@ -120,7 +120,7 @@ With Pix2Pix, you can input an image into the controlnet, and then "instruct" th
 Each of these models can be adjusted and combined with other ControlNet models to achieve different results, giving you even more control over your image generation process.
 
 
-## Using ControlNet
+### Using ControlNet
 
 To use ControlNet, you can simply select the desired model and adjust both the ControlNet and Pre-processor settings to achieve the desired result. You can also use multiple ControlNet models at the same time, allowing you to achieve even more complex effects or styles in your generated images.
 
@@ -132,3 +132,31 @@ Weight - Strength of the Controlnet model applied to the generation for the sect
 Start/End  - 0 represents the start of the generation, 1 represents the end. The Start/end setting controls what steps during the generation process have the ControlNet applied.
 
 Additionally, each ControlNet section can be expanded in order to manipulate settings for the image pre-processor that adjusts your uploaded image before using it in when you Invoke.
+
+
+## IP-Adapter
+
+[IP-Adapter](https://ip-adapter.github.io) is a tooling that allows for image prompt capabilities with text-to-image diffusion models. IP-Adapter works by analyzing the given image prompt to extract features, then passing those features to the UNet along with any other conditioning provided. 
+
+![IP-Adapter + T2I](https://github.com/tencent-ailab/IP-Adapter/raw/main/assets/demo/ip_adpter_plus_multi.jpg)
+
+![IP-Adapter + IMG2IMG](https://github.com/tencent-ailab/IP-Adapter/blob/main/assets/demo/image-to-image.jpg)
+
+#### Installation
+There are several ways to install IP-Adapter models with an existing InvokeAI installation:
+
+1. Through the command line interface launched from the invoke.sh / invoke.bat scripts, option [5] to download models.
+2. Through the Model Manager UI with models from the *Tools* section of [www.models.invoke.ai](www.models.invoke.ai). To do this, copy the repo ID from the desired model page, and paste it in the Add Model field of the model manager. **Note** Both the IP-Adapter and the Image Encoder must be installed for IP-Adapter to work. For example, the [SD 1.5 IP-Adapter](https://models.invoke.ai/InvokeAI/ip_adapter_plus_sd15) and [SD1.5 Image Encoder](https://models.invoke.ai/InvokeAI/ip_adapter_sd_image_encoder) must be installed to use IP-Adapter with SD1.5 based models.  
+3. **Advanced -- Not recommended ** Manually downloading the IP-Adapter and Image Encoder files - Image Encoder folders shouid be placed in the `models\any\clip_vision` folders. IP Adapter Model folders should be placed in the relevant `ip-adapter` folder of relevant base model folder of Invoke root directory. For example, for the SDXL IP-Adapter, files should be added to the `model/sdxl/ip_adapter/` folder. 
+
+#### Using IP-Adapter
+
+IP-Adapter can be used by navigating to the *Control Adapters* options and enabling IP-Adapter. 
+
+IP-Adapter requires an image to be used as the Image Prompt. It can also be used in conjunction with text prompts, Image-to-Image, Inpainting, Outpainting, ControlNets and LoRAs.
+
+
+Each IP-Adapter has two settings that are applied to the IP-Adapter:
+
+* Weight - Strength of the IP-Adapter model applied to the generation for the section, defined by start/end
+* Start/End  - 0 represents the start of the generation, 1 represents the end. The Start/end setting controls what steps during the generation process have the IP-Adapter applied.

docs/features/UTILITIES.md

@@ -0,0 +1,336 @@
+---
+title: Command-line Utilities
+---
+
+# :material-file-document: Utilities
+
+# Command-line Utilities
+
+InvokeAI comes with several scripts that are accessible via the
+command line. To access these commands, start the "developer's
+console" from the launcher (`invoke.bat` menu item [8]). Users who are
+familiar with Python can alternatively activate InvokeAI's virtual
+environment (typically, but not necessarily `invokeai/.venv`).
+
+In the developer's console, type the script's name to run it. To get a
+synopsis of what a utility does and the command-line arguments it
+accepts, pass it the `-h` argument, e.g.
+
+```bash
+invokeai-merge -h
+```
+## **invokeai-web**
+
+This script launches the web server and is effectively identical to
+selecting option [1] in the launcher. An advantage of launching the
+server from the command line is that you can override any setting
+configuration option in `invokeai.yaml` using like-named command-line
+arguments. For example, to temporarily change the size of the RAM
+cache to 7 GB, you can launch as follows:
+
+```bash
+invokeai-web --ram 7
+```
+
+## **invokeai-merge**
+
+This is the model merge script, the same as launcher option [4]. Call
+it with the `--gui` command-line argument to start the interactive
+console-based GUI. Alternatively, you can run it non-interactively
+using command-line arguments as illustrated in the example below which
+merges models named `stable-diffusion-1.5` and `inkdiffusion` into a new model named
+`my_new_model`:
+
+```bash
+invokeai-merge --force --base-model sd-1 --models stable-diffusion-1.5 inkdiffusion --merged_model_name my_new_model
+```
+
+## **invokeai-ti**
+
+This is the textual inversion training script that is run by launcher
+option [3]. Call it with `--gui` to run the interactive console-based
+front end. It can also be run non-interactively. It has about a
+zillion arguments, but a typical training session can be launched
+with:
+
+```bash
+invokeai-ti --model stable-diffusion-1.5 \
+            --placeholder_token 'jello' \
+            --learnable_property object \
+			--num_train_epochs 50 \
+			--train_data_dir /path/to/training/images \
+			--output_dir /path/to/trained/model
+```
+
+(Note that \\ is the Linux/Mac long-line continuation character. Use ^
+in Windows).
+
+## **invokeai-install**
+
+This is the console-based model install script that is run by launcher
+option [5]. If called without arguments, it will launch the
+interactive console-based interface. It can also be used
+non-interactively to list, add and remove models as shown by these
+examples:
+
+* This will download and install three models from CivitAI, HuggingFace,
+and local disk:
+
+```bash
+invokeai-install --add https://civitai.com/api/download/models/161302 ^
+                  gsdf/Counterfeit-V3.0  ^
+				  D:\Models\merge_model_two.safetensors
+```
+(Note that ^ is the Windows long-line continuation character. Use \\ on
+Linux/Mac).
+
+* This will list installed models of type `main`:
+
+```bash
+invokeai-model-install --list-models main
+```
+
+* This will delete the models named `voxel-ish` and `realisticVision`:
+
+```bash
+invokeai-model-install --delete voxel-ish realisticVision
+```
+
+## **invokeai-configure**
+
+This is the console-based configure script that ran when InvokeAI was
+first installed. You can run it again at any time to change the
+configuration, repair a broken install.
+
+Called without any arguments, `invokeai-configure` enters interactive
+mode with two screens. The first screen is a form that provides access
+to most of InvokeAI's configuration options. The second screen lets
+you download, add, and delete models interactively. When you exit the
+second screen, the script will add any missing "support models"
+needed for core functionality, and any selected "sd weights" which are
+the model checkpoint/diffusers files.
+
+This behavior can be changed via a series of command-line
+arguments. Here are some of the useful ones:
+
+* `invokeai-configure --skip-sd-weights --skip-support-models`
+This will run just the configuration part of the utility, skipping
+downloading of support models and stable diffusion weights.
+
+* `invokeai-configure --yes`
+This will run the configure script non-interactively. It will set the
+configuration options to their default values, install/repair support
+models, and download the "recommended" set of SD models.
+
+* `invokeai-configure --yes --default_only` 
+This will run the configure script non-interactively. In contrast to
+the previous command, it will only download the default SD model,
+Stable Diffusion v1.5
+
+* `invokeai-configure --yes --default_only --skip-sd-weights` 
+This is similar to the previous command, but will not download any
+SD models at all. It is usually used to repair a broken install.
+
+By default, `invokeai-configure` runs on the currently active InvokeAI
+root folder. To run it against a different root, pass it the `--root
+</path/to/root>` argument.
+
+Lastly, you can use `invokeai-configure` to create a working root
+directory entirely from scratch. Assuming you wish to make a root directory
+named `InvokeAI-New`, run this command:
+
+```bash
+invokeai-configure --root InvokeAI-New --yes --default_only
+```
+This will create a minimally functional root directory. You can now
+launch the web server against it with `invokeai-web --root InvokeAI-New`.
+
+## **invokeai-update**
+
+This is the interactive console-based script that is run by launcher
+menu item [9] to update to a new version of InvokeAI. It takes no
+command-line arguments.
+
+## **invokeai-metadata**
+
+This is a script which takes a list of InvokeAI-generated images and
+outputs their metadata in the same JSON format that you get from the
+`</>` button in the Web GUI. For example:
+
+```bash
+$ invokeai-metadata ffe2a115-b492-493c-afff-7679aa034b50.png
+ffe2a115-b492-493c-afff-7679aa034b50.png:
+{
+    "app_version": "3.1.0",
+    "cfg_scale": 8.0,
+    "clip_skip": 0,
+    "controlnets": [],
+    "generation_mode": "sdxl_txt2img",
+    "height": 1024,
+    "loras": [],
+    "model": {
+        "base_model": "sdxl",
+        "model_name": "stable-diffusion-xl-base-1.0",
+        "model_type": "main"
+    },
+    "negative_prompt": "",
+    "negative_style_prompt": "",
+    "positive_prompt": "military grade sushi dinner for shock troopers",
+    "positive_style_prompt": "",
+    "rand_device": "cpu",
+    "refiner_cfg_scale": 7.5,
+    "refiner_model": {
+        "base_model": "sdxl-refiner",
+        "model_name": "sd_xl_refiner_1.0",
+        "model_type": "main"
+    },
+    "refiner_negative_aesthetic_score": 2.5,
+    "refiner_positive_aesthetic_score": 6.0,
+    "refiner_scheduler": "euler",
+    "refiner_start": 0.8,
+    "refiner_steps": 20,
+    "scheduler": "euler",
+    "seed": 387129902,
+    "steps": 25,
+    "width": 1024
+}
+```
+
+You may list multiple files on the command line.
+
+## **invokeai-import-images**
+
+InvokeAI uses a database to store information about images it
+generated, and just copying the image files from one InvokeAI root
+directory to another does not automatically import those images into
+the destination's gallery. This script allows you to bulk import
+images generated by one instance of InvokeAI into a gallery maintained
+by another. It also works on images generated by older versions of
+InvokeAI, going way back to version 1.
+
+This script has an interactive mode only. The following example shows
+it in action:
+
+```bash
+$ invokeai-import-images
+===============================================================================
+This script will import images generated by earlier versions of
+InvokeAI into the currently installed root directory:
+   /home/XXXX/invokeai-main
+If this is not what you want to do, type ctrl-C now to cancel.
+===============================================================================
+= Configuration & Settings
+Found invokeai.yaml file at /home/XXXX/invokeai-main/invokeai.yaml:
+  Database : /home/XXXX/invokeai-main/databases/invokeai.db
+  Outputs  : /home/XXXX/invokeai-main/outputs/images
+ 
+Use these paths for import (yes) or choose different ones (no) [Yn]:
+Inputs: Specify absolute path containing InvokeAI .png images to import: /home/XXXX/invokeai-2.3/outputs/images/
+Include files from subfolders recursively [yN]?
+
+Options for board selection for imported images:
+1) Select an existing board name. (found 4)
+2) Specify a board name to create/add to.
+3) Create/add to board named 'IMPORT'.
+4) Create/add to board named 'IMPORT' with the current datetime string appended (.e.g IMPORT_20230919T203519Z).
+5) Create/add to board named 'IMPORT' with a the original file app_version appended (.e.g IMPORT_2.2.5).
+Specify desired board option: 3
+
+===============================================================================
+= Import Settings Confirmation
+
+Database File Path               : /home/XXXX/invokeai-main/databases/invokeai.db
+Outputs/Images Directory         : /home/XXXX/invokeai-main/outputs/images
+Import Image Source Directory    : /home/XXXX/invokeai-2.3/outputs/images/
+  Recurse Source SubDirectories  : No
+Count of .png file(s) found      : 5785
+Board name option specified      : IMPORT
+Database backup will be taken at : /home/XXXX/invokeai-main/databases/backup
+
+Notes about the import process:
+- Source image files will not be modified, only copied to the outputs directory.
+- If the same file name already exists in the destination, the file will be skipped.
+- If the same file name already has a record in the database, the file will be skipped.
+- Invoke AI metadata tags will be updated/written into the imported copy only.
+- On the imported copy, only Invoke AI known tags (latest and legacy) will be retained (dream, sd-metadata, invokeai, invokeai_metadata)
+- A property 'imported_app_version' will be added to metadata that can be viewed in the UI's metadata viewer.
+- The new 3.x InvokeAI outputs folder structure is flat so recursively found source imges will all be placed into the single outputs/images folder.
+ 
+Do you wish to continue with the import [Yn] ?
+
+Making DB Backup at /home/lstein/invokeai-main/databases/backup/backup-20230919T203519Z-invokeai.db...Done!
+
+===============================================================================
+Importing /home/XXXX/invokeai-2.3/outputs/images/17d09907-297d-4db3-a18a-60b337feac66.png
+... (5785 more lines) ...
+===============================================================================
+= Import Complete - Elpased Time: 0.28 second(s)
+
+Source File(s)                          : 5785
+Total Imported                          : 5783
+Skipped b/c file already exists on disk : 1
+Skipped b/c file already exists in db   : 0
+Errors during import                    : 1
+```
+## **invokeai-db-maintenance**
+
+This script helps maintain the integrity of your InvokeAI database by
+finding and fixing three problems that can arise over time:
+
+1. An image was manually deleted from the outputs directory, leaving a
+   dangling image record in the InvokeAI database. This will cause a
+   black image to appear in the gallery. This is an "orphaned database
+   image record." The script can fix this by running a "clean"
+   operation on the database, removing the orphaned entries.
+   
+2. An image is present in the outputs directory but there is no
+   corresponding entry in the database. This can happen when the image
+   is added manually to the outputs directory, or if a crash occurred
+   after the image was generated but before the database was
+   completely updated. The symptom is that the image is present in the
+   outputs folder but doesn't appear in the InvokeAI gallery. This is
+   called an "orphaned image file." The script can fix this problem by
+   running an "archive" operation in which orphaned files are moved
+   into a directory named `outputs/images-archive`. If you wish, you
+   can then run `invokeai-image-import` to reimport these images back
+   into the database.
+   
+3. The thumbnail for an image is missing, again causing a black
+   gallery thumbnail. This is fixed by running the "thumbnaiils"
+   operation, which simply regenerates and re-registers the missing
+   thumbnail.
+   
+You can find and fix all three of these problems in a single go by
+executing this command:
+
+```bash
+invokeai-db-maintenance --operation all
+```
+
+Or you can run just the clean and thumbnail operations like this:
+
+```bash
+invokeai-db-maintenance -operation clean, thumbnail
+```
+
+If called without any arguments, the script will ask you which
+operations you wish to perform.
+
+## **invokeai-migrate3**
+
+This script will migrate settings and models (but not images!) from an
+InvokeAI v2.3 root folder to an InvokeAI 3.X folder. Call it with the
+source and destination root folders like this:
+
+```bash
+invokeai-migrate3 --from ~/invokeai-2.3 --to invokeai-3.1.1
+```
+
+Both directories must previously have been properly created and
+initialized by `invokeai-configure`. If you wish to migrate the images
+contained in the older root as well, you can use the
+`invokeai-image-migrate` script described earlier.
+
+---
+
+Copyright (c) 2023, Lincoln Stein and the InvokeAI Development Team

docs/features/index.md

@@ -51,6 +51,9 @@ Prevent InvokeAI from displaying unwanted racy images.
 ### * [Controlling Logging](LOGGING.md)
 Control how InvokeAI logs status messages.
 
+### * [Command-line Utilities](UTILITIES.md)
+A list of the command-line utilities available with InvokeAI.
+
 <!-- OUT OF DATE
 ### * [Miscellaneous](OTHER.md)
 Run InvokeAI on Google Colab, generate images with repeating patterns,

docs/index.md

@@ -147,6 +147,7 @@ Mac and Linux machines, and runs on GPU cards with as little as 4 GB of RAM.
 
 ### InvokeAI Configuration
 - [Guide to InvokeAI Runtime Settings](features/CONFIGURATION.md)
+- [Database Maintenance and other Command Line Utilities](features/UTILITIES.md)
 
 ## :octicons-log-16: Important Changes Since Version 2.3
 

docs/installation/020_INSTALL_MANUAL.md

@@ -256,6 +256,10 @@ manager, please follow these steps:
         *highly recommended** if your virtual environment is located outside of
         your runtime directory.
 
+    !!! tip
+
+        On linux, it is recommended to run invokeai with the following env var: `MALLOC_MMAP_THRESHOLD_=1048576`. For example: `MALLOC_MMAP_THRESHOLD_=1048576 invokeai --web`. This helps to prevent memory fragmentation that can lead to memory accumulation over time. This env var is set automatically when running via `invoke.sh`.
+
 10.  Render away!
 
     Browse the [features](../features/index.md) section to learn about all the
@@ -296,8 +300,18 @@ code for InvokeAI. For this to work, you will need to install the
 on your system, please see the [Git Installation
 Guide](https://github.com/git-guides/install-git)
 
+You will also need to install the [frontend development toolchain](https://github.com/invoke-ai/InvokeAI/blob/main/docs/contributing/contribution_guides/contributingToFrontend.md).
+
+If you have a "normal" installation, you should create a totally separate virtual environment for the git-based installation, else the two may interfere.
+
+> **Why do I need the frontend toolchain**?
+>
+> The InvokeAI project uses trunk-based development. That means our `main` branch is the development branch, and releases are tags on that branch. Because development is very active, we don't keep an updated build of the UI in `main` - we only build it for production releases.
+>
+> That means that between releases, to have a functioning application when running directly from the repo, you will need to run the UI in dev mode or build it regularly (any time the UI code changes).
+
 1. Create a fork of the InvokeAI repository through the GitHub UI or [this link](https://github.com/invoke-ai/InvokeAI/fork) 
-1. From the command line, run this command:
+2. From the command line, run this command:
    ```bash
    git clone https://github.com/<your_github_username>/InvokeAI.git
    ```
@@ -305,10 +319,10 @@ Guide](https://github.com/git-guides/install-git)
     This will create a directory named `InvokeAI` and populate it with the
     full source code from your fork of the InvokeAI repository.
 
-2. Activate the InvokeAI virtual environment as per step (4) of the manual
+3. Activate the InvokeAI virtual environment as per step (4) of the manual
 installation protocol (important!)
 
-3. Enter the InvokeAI repository directory and run one of these
+4. Enter the InvokeAI repository directory and run one of these
    commands, based on your GPU:
 
     === "CUDA (NVidia)"
@@ -334,11 +348,15 @@ installation protocol (important!)
     Be sure to pass `-e` (for an editable install) and don't forget the
     dot ("."). It is part of the command.
 
-    You can now run `invokeai` and its related commands. The code will be
+5.  Install the [frontend toolchain](https://github.com/invoke-ai/InvokeAI/blob/main/docs/contributing/contribution_guides/contributingToFrontend.md) and do a production build of the UI as described.
+
+6.  You can now run `invokeai` and its related commands. The code will be
     read from the repository, so that you can edit the .py source files
     and watch the code's behavior change.
 
-4.  If you wish to contribute to the InvokeAI project, you are
+    When you pull in new changes to the repo, be sure to re-build the UI.
+
+7.  If you wish to contribute to the InvokeAI project, you are
     encouraged to establish a GitHub account and "fork"
     https://github.com/invoke-ai/InvokeAI into your own copy of the
     repository. You can then use GitHub functions to create and submit

docs/installation/050_INSTALLING_MODELS.md

@@ -171,3 +171,16 @@ subfolders and organize them as you wish.
 
 The location of the autoimport directories are controlled by settings
 in `invokeai.yaml`. See [Configuration](../features/CONFIGURATION.md).
+
+### Installing models that live in HuggingFace subfolders
+
+On rare occasions you may need to install a diffusers-style model that
+lives in a subfolder of a HuggingFace repo id. In this event, simply
+add ":_subfolder-name_" to the end of the repo id. For example, if the
+repo id is "monster-labs/control_v1p_sd15_qrcode_monster" and the model
+you wish to fetch lives in a subfolder named "v2", then the repo id to
+pass to the various model installers should be 
+
+```
+monster-labs/control_v1p_sd15_qrcode_monster:v2
+```

docs/nodes/NODES.md

@@ -4,12 +4,12 @@ The workflow editor is a blank canvas allowing for the use of individual functio
 
 If you're not familiar with Diffusion, take a look at our [Diffusion Overview.](../help/diffusion.md) Understanding how diffusion works will enable you to more easily use the Workflow Editor and build workflows to suit your needs.
 
-## UI Features
+## Features
 
 ### Linear View
 The Workflow Editor allows you to create a UI for your workflow, to make it easier to iterate on your generations. 
 
-To add an input to the Linear UI, right click on the input and select "Add to Linear View".
+To add an input to the Linear UI, right click on the input label and select "Add to Linear View".
 
 The Linear UI View will also be part of the saved workflow, allowing you share workflows and enable other to use them, regardless of complexity. 
 
@@ -25,6 +25,10 @@ Any node or input field can be renamed in the workflow editor. If the input fiel
 * Backspace/Delete to delete a node
 * Shift+Click to drag and select multiple nodes 
 
+### Node Caching 
+
+Nodes have a "Use Cache" option in their footer. This allows for performance improvements by using the previously cached values during the workflow processing. 
+
 
 ## Important Concepts 
 

docs/nodes/communityNodes.md

@@ -8,19 +8,21 @@ To download a node, simply download the `.py` node file from the link and add it
 
 To use a community workflow, download the the `.json` node graph file and load it into Invoke AI via the **Load Workflow** button in the Workflow Editor. 
 
-## Community Nodes
+--------------------------------
 
-### FaceTools
+--------------------------------
+### Make 3D
 
-**Description:** FaceTools is a collection of nodes created to manipulate faces as you would in Unified Canvas. It includes FaceMask, FaceOff, and FacePlace. FaceMask autodetects a face in the image using MediaPipe and creates a mask from it. FaceOff similarly detects a face, then takes the face off of the image by adding a square bounding box around it and cropping/scaling it. FacePlace puts the bounded face image from FaceOff back onto the original image. Using these nodes with other inpainting node(s), you can put new faces on existing things, put new things around existing faces, and work closer with a face as a bounded image. Additionally, you can supply X and Y offset values to scale/change the shape of the mask for finer control on FaceMask and FaceOff. See GitHub repository below for usage examples.
+**Description:** Create compelling 3D stereo images from 2D originals.
 
-**Node Link:** https://github.com/ymgenesis/FaceTools/
+**Node Link:** [https://gitlab.com/srcrr/shift3d/-/raw/main/make3d.py](https://gitlab.com/srcrr/shift3d)
 
-**FaceMask Output Examples** 
+**Example Node Graph:**  https://gitlab.com/srcrr/shift3d/-/raw/main/example-workflow.json?ref_type=heads&inline=false
 
-![5cc8abce-53b0-487a-b891-3bf94dcc8960](https://github.com/invoke-ai/InvokeAI/assets/25252829/43f36d24-1429-4ab1-bd06-a4bedfe0955e)
-![b920b710-1882-49a0-8d02-82dff2cca907](https://github.com/invoke-ai/InvokeAI/assets/25252829/7660c1ed-bf7d-4d0a-947f-1fc1679557ba)
-![71a91805-fda5-481c-b380-264665703133](https://github.com/invoke-ai/InvokeAI/assets/25252829/f8f6a2ee-2b68-4482-87da-b90221d5c3e2)
+**Output Examples** 
+
+![Painting of a cozy delapidated house](https://gitlab.com/srcrr/shift3d/-/raw/main/example-1.png){: style="height:512px;width:512px"}
+![Photo of cute puppies](https://gitlab.com/srcrr/shift3d/-/raw/main/example-2.png){: style="height:512px;width:512px"}
 
 --------------------------------
 ### Ideal Size
@@ -43,6 +45,52 @@ To use a community workflow, download the the `.json` node graph file and load i
 
 **Node Link:** https://github.com/JPPhoto/image-picker-node
 
+--------------------------------
+### Thresholding
+
+**Description:** This node generates masks for highlights, midtones, and shadows given an input image. You can optionally specify a blur for the lookup table used in making those masks from the source image.
+
+**Node Link:** https://github.com/JPPhoto/thresholding-node
+
+**Examples**
+
+Input:
+
+![image](https://github.com/invoke-ai/InvokeAI/assets/34005131/c88ada13-fb3d-484c-a4fe-947b44712632){: style="height:512px;width:512px"}
+
+Highlights/Midtones/Shadows:
+
+<img src="https://github.com/invoke-ai/InvokeAI/assets/34005131/727021c1-36ff-4ec8-90c8-105e00de986d" style="width: 30%" />
+<img src="https://github.com/invoke-ai/InvokeAI/assets/34005131/0b721bfc-f051-404e-b905-2f16b824ddfe" style="width: 30%" />
+<img src="https://github.com/invoke-ai/InvokeAI/assets/34005131/04c1297f-1c88-42b6-a7df-dd090b976286" style="width: 30%" />
+
+Highlights/Midtones/Shadows (with LUT blur enabled):
+
+<img src="https://github.com/invoke-ai/InvokeAI/assets/34005131/19aa718a-70c1-4668-8169-d68f4bd13771" style="width: 30%" />
+<img src="https://github.com/invoke-ai/InvokeAI/assets/34005131/0a440e43-697f-4d17-82ee-f287467df0a5" style="width: 30%" />
+<img src="https://github.com/invoke-ai/InvokeAI/assets/34005131/0701fd0f-2ca7-4fe2-8613-2b52547bafce" style="width: 30%" />
+
+--------------------------------
+### Halftone
+
+**Description**: Halftone converts the source image to grayscale and then performs halftoning. CMYK Halftone converts the image to CMYK and applies a per-channel halftoning to make the source image look like a magazine or newspaper. For both nodes, you can specify angles and halftone dot spacing.
+
+**Node Link:** https://github.com/JPPhoto/halftone-node
+
+**Example**
+
+Input:
+
+![image](https://github.com/invoke-ai/InvokeAI/assets/34005131/fd5efb9f-4355-4409-a1c2-c1ca99e0cab4){: style="height:512px;width:512px"}
+
+Halftone Output:
+
+![image](https://github.com/invoke-ai/InvokeAI/assets/34005131/7e606f29-e68f-4d46-b3d5-97f799a4ec2f){: style="height:512px;width:512px"}
+
+CMYK Halftone Output:
+
+![image](https://github.com/invoke-ai/InvokeAI/assets/34005131/c59c578f-db8e-4d66-8c66-2851752d75ea){: style="height:512px;width:512px"}
+
 --------------------------------
 ### Retroize
 
@@ -77,7 +125,7 @@ Generated Prompt: An enchanted weapon will be usable by any character regardless
 **Example Node Graph:**  https://github.com/helix4u/load_video_frame/blob/main/Example_Workflow.json
 
 **Output Example:** 
-=======
+
 ![Example animation](https://github.com/helix4u/load_video_frame/blob/main/testmp4_embed_converted.gif)
 [Full mp4 of Example Output test.mp4](https://github.com/helix4u/load_video_frame/blob/main/test.mp4)
 
@@ -121,18 +169,6 @@ To be imported, an .obj must use triangulated meshes, so make sure to enable tha
 **Example Usage:**
 ![depth from obj usage graph](https://raw.githubusercontent.com/dwringer/depth-from-obj-node/main/depth_from_obj_usage.jpg)
 
---------------------------------
-### Enhance Image (simple adjustments)
-
-**Description:** Boost or reduce color saturation, contrast, brightness, sharpness, or invert colors of any image at any stage with this simple wrapper for pillow [PIL]'s ImageEnhance module.
-
-Color inversion is toggled with a simple switch, while each of the four enhancer modes are activated by entering a value other than 1 in each corresponding input field. Values less than 1 will reduce the corresponding property, while values greater than 1 will enhance it.
-
-**Node Link:** https://github.com/dwringer/image-enhance-node
-
-**Example Usage:**
-![enhance image usage graph](https://raw.githubusercontent.com/dwringer/image-enhance-node/main/image_enhance_usage.jpg)
-
 --------------------------------
 ### Generative Grammar-Based Prompt Nodes
 
@@ -153,16 +189,28 @@ This includes 3 Nodes:
 
 **Description:** This is a pack of nodes for composing masks and images, including a simple text mask creator and both image and latent offset nodes. The offsets wrap around, so these can be used in conjunction with the Seamless node to progressively generate centered on different parts of the seamless tiling.
 
-This includes 4 Nodes:
-- *Text Mask (simple 2D)* - create and position a white on black (or black on white) line of text using any font locally available to Invoke.
+This includes 15 Nodes:
+
+- *Adjust Image Hue Plus* - Rotate the hue of an image in one of several different color spaces.
+- *Blend Latents/Noise (Masked)* - Use a mask to blend part of one latents tensor [including Noise outputs] into another. Can be used to "renoise" sections during a multi-stage [masked] denoising process.
+- *Enhance Image* - Boost or reduce color saturation, contrast, brightness, sharpness, or invert colors of any image at any stage with this simple wrapper for pillow [PIL]'s ImageEnhance module.
+- *Equivalent Achromatic Lightness* - Calculates image lightness accounting for Helmholtz-Kohlrausch effect based on a method described by High, Green, and Nussbaum (2023).
+- *Text to Mask (Clipseg)* - Input a prompt and an image to generate a mask representing areas of the image matched by the prompt.
+- *Text to Mask Advanced (Clipseg)* - Output up to four prompt masks combined with logical "and", logical "or", or as separate channels of an RGBA image.
+- *Image Layer Blend* - Perform a layered blend of two images using alpha compositing. Opacity of top layer is selectable, with optional mask and several different blend modes/color spaces.
 - *Image Compositor* - Take a subject from an image with a flat backdrop and layer it on another image using a chroma key or flood select background removal.
+- *Image Dilate or Erode* - Dilate or expand a mask (or any image!). This is equivalent to an expand/contract operation.
+- *Image Value Thresholds* - Clip an image to pure black/white beyond specified thresholds.
 - *Offset Latents* - Offset a latents tensor in the vertical and/or horizontal dimensions, wrapping it around.
 - *Offset Image* - Offset an image in the vertical and/or horizontal dimensions, wrapping it around.
+- *Rotate/Flip Image* - Rotate an image in degrees clockwise/counterclockwise about its center, optionally resizing the image boundaries to fit, or flipping it about the vertical and/or horizontal axes.
+- *Shadows/Highlights/Midtones* - Extract three masks (with adjustable hard or soft thresholds) representing shadows, midtones, and highlights regions of an image.
+- *Text Mask (simple 2D)* - create and position a white on black (or black on white) line of text using any font locally available to Invoke.
 
 **Node Link:** https://github.com/dwringer/composition-nodes
 
-**Example Usage:**
-![composition nodes usage graph](https://raw.githubusercontent.com/dwringer/composition-nodes/main/composition_nodes_usage.jpg)
+**Nodes and Output Examples:**
+![composition nodes usage graph](https://raw.githubusercontent.com/dwringer/composition-nodes/main/composition_pack_overview.jpg)
 
 --------------------------------
 ### Size Stepper Nodes
@@ -196,6 +244,70 @@ Results after using the depth controlnet
 
 --------------------------------
 
+### Prompt Tools 
+
+**Description:** A set of InvokeAI nodes that add general prompt manipulation tools.  These where written to accompany the PromptsFromFile node and other prompt generation nodes.
+
+1. PromptJoin - Joins to prompts into one.
+2. PromptReplace - performs a search and replace on a prompt. With the option of using regex.
+3. PromptSplitNeg - splits a prompt into positive and negative using the old V2 method of [] for negative.
+4. PromptToFile - saves a prompt or collection of prompts to a file. one per line. There is an append/overwrite option.
+5. PTFieldsCollect - Converts image generation fields into a Json format string that can be passed to Prompt to file. 
+6. PTFieldsExpand - Takes Json string and converts it to individual generation parameters This can be fed from the Prompts to file node.
+7. PromptJoinThree -  Joins 3 prompt together.
+8. PromptStrength - This take a string and float and outputs another string in the format of (string)strength like the weighted format of compel. 
+9. PromptStrengthCombine - This takes a collection of prompt strength strings and outputs a string in the .and() or .blend() format that can be fed into a proper prompt node.
+
+See full docs here: https://github.com/skunkworxdark/Prompt-tools-nodes/edit/main/README.md
+
+**Node Link:** https://github.com/skunkworxdark/Prompt-tools-nodes
+
+--------------------------------
+
+### XY Image to Grid and Images to Grids nodes
+
+**Description:** Image to grid nodes and supporting tools.
+
+1. "Images To Grids" node - Takes a collection of images and creates a grid(s) of images. If there are more images than the size of a single grid then mutilple grids will be created until it runs out of images.
+2. "XYImage To Grid" node - Converts a collection of XYImages into a labeled Grid of images.  The XYImages collection has to be built using the supporoting nodes. See example node setups for more details.
+   
+
+See full docs here: https://github.com/skunkworxdark/XYGrid_nodes/edit/main/README.md
+
+**Node Link:** https://github.com/skunkworxdark/XYGrid_nodes
+
+--------------------------------
+
+### Image to Character Art Image Node's
+
+**Description:** Group of nodes to convert an input image into ascii/unicode art Image
+
+**Node Link:** https://github.com/mickr777/imagetoasciiimage
+
+**Output Examples**
+
+<img src="https://github.com/invoke-ai/InvokeAI/assets/115216705/8e061fcc-9a2c-4fa9-bcc7-c0f7b01e9056" width="300" />
+<img src="https://github.com/mickr777/imagetoasciiimage/assets/115216705/3c4990eb-2f42-46b9-90f9-0088b939dc6a" width="300" /></br>
+<img src="https://github.com/mickr777/imagetoasciiimage/assets/115216705/fee7f800-a4a8-41e2-a66b-c66e4343307e" width="300" />
+<img src="https://github.com/mickr777/imagetoasciiimage/assets/115216705/1d9c1003-a45f-45c2-aac7-46470bb89330" width="300" />
+
+--------------------------------
+
+### Grid to Gif
+
+**Description:** One node that turns a grid image into an image colletion, one node that turns an image collection into a gif
+
+**Node Link:** https://github.com/mildmisery/invokeai-GridToGifNode/blob/main/GridToGif.py
+
+**Example Node Graph:**  https://github.com/mildmisery/invokeai-GridToGifNode/blob/main/Grid%20to%20Gif%20Example%20Workflow.json
+
+**Output Examples** 
+
+<img src="https://raw.githubusercontent.com/mildmisery/invokeai-GridToGifNode/main/input.png" width="300" />
+<img src="https://raw.githubusercontent.com/mildmisery/invokeai-GridToGifNode/main/output.gif" width="300" />
+
+--------------------------------
+
 ### Example Node Template
 
 **Description:** This node allows you to do super cool things with InvokeAI.

docs/nodes/defaultNodes.md

@@ -1,6 +1,6 @@
 # List of Default Nodes
 
-The table below contains a list of the default nodes shipped with InvokeAI and their descriptions. 
+The table below contains a list of the default nodes shipped with InvokeAI and their descriptions.
 
 | Node <img width=160 align="right"> | Function                                                                              |
 |: ---------------------------------- | :--------------------------------------------------------------------------------------|
@@ -17,11 +17,12 @@ The table below contains a list of the default nodes shipped with InvokeAI and t
 |Conditioning Primitive 			| A conditioning tensor primitive value|
 |Content Shuffle Processor 			| Applies content shuffle processing to image|
 |ControlNet 			| Collects ControlNet info to pass to other nodes|
-|OpenCV Inpaint 			| Simple inpaint using opencv.|
 |Denoise Latents 			| Denoises noisy latents to decodable images|
 |Divide Integers 			| Divides two numbers|
 |Dynamic Prompt 			| Parses a prompt using adieyal/dynamicprompts' random or combinatorial generator|
-|Upscale (RealESRGAN) 			| Upscales an image using RealESRGAN.|
+|[FaceMask](./detailedNodes/faceTools.md#facemask) | Generates masks for faces in an image to use with Inpainting|
+|[FaceIdentifier](./detailedNodes/faceTools.md#faceidentifier)             | Identifies and labels faces in an image|
+|[FaceOff](./detailedNodes/faceTools.md#faceoff)             | Creates a new image that is a scaled bounding box with a mask on the face for Inpainting|
 |Float Math             | Perform basic math operations on two floats|
 |Float Primitive Collection 			| A collection of float primitive values|
 |Float Primitive 			| A float primitive value|
@@ -76,6 +77,7 @@ The table below contains a list of the default nodes shipped with InvokeAI and t
 |ONNX Prompt (Raw) 			| A node to process inputs and produce outputs. May use dependency injection in __init__ to receive providers.|
 |ONNX Text to Latents 			| Generates latents from conditionings.|
 |ONNX Model Loader 			| Loads a main model, outputting its submodels.|
+|OpenCV Inpaint 			| Simple inpaint using opencv.|
 |Openpose Processor 			| Applies Openpose processing to image|
 |PIDI Processor 			| Applies PIDI processing to image|
 |Prompts from File 			| Loads prompts from a text file|
@@ -97,5 +99,6 @@ The table below contains a list of the default nodes shipped with InvokeAI and t
 |String Primitive 			| A string primitive value|
 |Subtract Integers 			| Subtracts two numbers|
 |Tile Resample Processor 			| Tile resampler processor|
+|Upscale (RealESRGAN) 			| Upscales an image using RealESRGAN.|
 |VAE Loader 			| Loads a VAE model, outputting a VaeLoaderOutput|
 |Zoe (Depth) Processor 			| Applies Zoe depth processing to image|

docs/nodes/detailedNodes/faceTools.md

@@ -0,0 +1,154 @@
+# Face Nodes
+
+## FaceOff
+
+FaceOff mimics a user finding a face in an image and resizing the bounding box
+around the head in Canvas.
+
+Enter a face ID (found with FaceIdentifier) to choose which face to mask.
+
+Just as you would add more context inside the bounding box by making it larger
+in Canvas, the node gives you a padding input (in pixels) which will
+simultaneously add more context, and increase the resolution of the bounding box
+so the face remains the same size inside it.
+
+The "Minimum Confidence" input defaults to 0.5 (50%), and represents a pass/fail
+threshold a detected face must reach for it to be processed. Lowering this value
+may help if detection is failing. If the detected masks are imperfect and stray
+too far outside/inside of faces, the node gives you X & Y offsets to shrink/grow
+the masks by a multiplier.
+
+FaceOff will output the face in a bounded image, taking the face off of the
+original image for input into any node that accepts image inputs. The node also
+outputs a face mask with the dimensions of the bounded image. The X & Y outputs
+are for connecting to the X & Y inputs of the Paste Image node, which will place
+the bounded image back on the original image using these coordinates.
+
+###### Inputs/Outputs
+
+| Input              | Description                                                                                                                                                                              |
+| ------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Image              | Image for face detection                                                                                                                                                                 |
+| Face ID            | The face ID to process, numbered from 0. Multiple faces not supported. Find a face's ID with FaceIdentifier node.                                                                        |
+| Minimum Confidence | Minimum confidence for face detection (lower if detection is failing)                                                                                                                    |
+| X Offset           | X-axis offset of the mask                                                                                                                                                                |
+| Y Offset           | Y-axis offset of the mask                                                                                                                                                                |
+| Padding            | All-axis padding around the mask in pixels                                                                                                                                               |
+| Chunk              | Chunk (or divide) the image into sections to greatly improve face detection success. Defaults to off, but will activate if no faces are detected normally. Activate to chunk by default. |
+
+| Output        | Description                                      |
+| ------------- | ------------------------------------------------ |
+| Bounded Image | Original image bound, cropped, and resized       |
+| Width         | The width of the bounded image in pixels         |
+| Height        | The height of the bounded image in pixels        |
+| Mask          | The output mask                                  |
+| X             | The x coordinate of the bounding box's left side |
+| Y             | The y coordinate of the bounding box's top side  |
+
+## FaceMask
+
+FaceMask mimics a user drawing masks on faces in an image in Canvas.
+
+The "Face IDs" input allows the user to select specific faces to be masked.
+Leave empty to detect and mask all faces, or a comma-separated list for a
+specific combination of faces (ex: `1,2,4`). A single integer will detect and
+mask that specific face. Find face IDs with the FaceIdentifier node.
+
+The "Minimum Confidence" input defaults to 0.5 (50%), and represents a pass/fail
+threshold a detected face must reach for it to be processed. Lowering this value
+may help if detection is failing.
+
+If the detected masks are imperfect and stray too far outside/inside of faces,
+the node gives you X & Y offsets to shrink/grow the masks by a multiplier. All
+masks shrink/grow together by the X & Y offset values.
+
+By default, masks are created to change faces. When masks are inverted, they
+change surrounding areas, protecting faces.
+
+###### Inputs/Outputs
+
+| Input              | Description                                                                                                                                                                              |
+| ------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Image              | Image for face detection                                                                                                                                                                 |
+| Face IDs           | Comma-separated list of face ids to mask eg '0,2,7'. Numbered from 0. Leave empty to mask all. Find face IDs with FaceIdentifier node.                                                   |
+| Minimum Confidence | Minimum confidence for face detection (lower if detection is failing)                                                                                                                    |
+| X Offset           | X-axis offset of the mask                                                                                                                                                                |
+| Y Offset           | Y-axis offset of the mask                                                                                                                                                                |
+| Chunk              | Chunk (or divide) the image into sections to greatly improve face detection success. Defaults to off, but will activate if no faces are detected normally. Activate to chunk by default. |
+| Invert Mask        | Toggle to invert the face mask                                                                                                                                                           |
+
+| Output | Description                       |
+| ------ | --------------------------------- |
+| Image  | The original image                |
+| Width  | The width of the image in pixels  |
+| Height | The height of the image in pixels |
+| Mask   | The output face mask              |
+
+## FaceIdentifier
+
+FaceIdentifier outputs an image with detected face IDs printed in white numbers
+onto each face.
+
+Face IDs can then be used in FaceMask and FaceOff to selectively mask all, a
+specific combination, or single faces.
+
+The FaceIdentifier output image is generated for user reference, and isn't meant
+to be passed on to other image-processing nodes.
+
+The "Minimum Confidence" input defaults to 0.5 (50%), and represents a pass/fail
+threshold a detected face must reach for it to be processed. Lowering this value
+may help if detection is failing. If an image is changed in the slightest, run
+it through FaceIdentifier again to get updated FaceIDs.
+
+###### Inputs/Outputs
+
+| Input              | Description                                                                                                                                                                              |
+| ------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Image              | Image for face detection                                                                                                                                                                 |
+| Minimum Confidence | Minimum confidence for face detection (lower if detection is failing)                                                                                                                    |
+| Chunk              | Chunk (or divide) the image into sections to greatly improve face detection success. Defaults to off, but will activate if no faces are detected normally. Activate to chunk by default. |
+
+| Output | Description                                                                                      |
+| ------ | ------------------------------------------------------------------------------------------------ |
+| Image  | The original image with small face ID numbers printed in white onto each face for user reference |
+| Width  | The width of the original image in pixels                                                        |
+| Height | The height of the original image in pixels                                                       |
+
+## Tips
+
+- If not all target faces are being detected, activate Chunk to bypass full
+  image face detection and greatly improve detection success.
+- Final results will vary between full-image detection and chunking for faces
+  that are detectable by both due to the nature of the process. Try either to
+  your taste.
+- Be sure Minimum Confidence is set the same when using FaceIdentifier with
+  FaceOff/FaceMask.
+- For FaceOff, use the color correction node before faceplace to correct edges
+  being noticeable in the final image (see example screenshot).
+- Non-inpainting models may struggle to paint/generate correctly around faces.
+- If your face won't change the way you want it to no matter what you change,
+  consider that the change you're trying to make is too much at that resolution.
+  For example, if an image is only 512x768 total, the face might only be 128x128
+  or 256x256, much smaller than the 512x512 your SD1.5 model was probably
+  trained on. Try increasing the resolution of the image by upscaling or
+  resizing, add padding to increase the bounding box's resolution, or use an
+  image where the face takes up more pixels.
+- If the resulting face seems out of place pasted back on the original image
+  (ie. too large, not proportional), add more padding on the FaceOff node to
+  give inpainting more context. Context and good prompting are important to
+  keeping things proportional.
+- If you find the mask is too big/small and going too far outside/inside the
+  area you want to affect, adjust the x & y offsets to shrink/grow the mask area
+- Use a higher denoise start value to resemble aspects of the original face or
+  surroundings. Denoise start = 0 & denoise end = 1 will make something new,
+  while denoise start = 0.50 & denoise end = 1 will be 50% old and 50% new.
+- mediapipe isn't good at detecting faces with lots of face paint, hair covering
+  the face, etc. Anything that obstructs the face will likely result in no faces
+  being detected.
+- If you find your face isn't being detected, try lowering the minimum
+  confidence value from 0.5. This could result in false positives, however
+  (random areas being detected as faces and masked).
+- After altering an image and wanting to process a different face in the newly
+  altered image, run the altered image through FaceIdentifier again to see the
+  new Face IDs. MediaPipe will most likely detect faces in a different order
+  after an image has been changed in the slightest.

docs/nodes/exampleWorkflows.md

@@ -9,5 +9,6 @@ If you're interested in finding more workflows, checkout the [#share-your-workfl
 * [SD1.5 / SD2 Text to Image](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/Text_to_Image.json)
 * [SDXL Text to Image](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/SDXL_Text_to_Image.json)
 * [SDXL (with Refiner) Text to Image](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/SDXL_Text_to_Image.json) 
-* [Tiled Upscaling with ControlNet](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/ESRGAN_img2img_upscale w_Canny_ControlNet.json)ß
-
+* [Tiled Upscaling with ControlNet](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/ESRGAN_img2img_upscale w_Canny_ControlNet.json)
+* [FaceMask](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/FaceMask.json)
+* [FaceOff with 2x Face Scaling](https://github.com/invoke-ai/InvokeAI/blob/main/docs/workflows/FaceOff_FaceScale2x.json)

installer/lib/installer.py

@@ -332,6 +332,7 @@ class InvokeAiInstance:
         Configure the InvokeAI runtime directory
         """
 
+        auto_install = False
         # set sys.argv to a consistent state
         new_argv = [sys.argv[0]]
         for i in range(1, len(sys.argv)):
@@ -340,13 +341,17 @@ class InvokeAiInstance:
                 new_argv.append(el)
                 new_argv.append(sys.argv[i + 1])
             elif el in ["-y", "--yes", "--yes-to-all"]:
-                new_argv.append(el)
+                auto_install = True
         sys.argv = new_argv
 
+        import messages
         import requests  # to catch download exceptions
-        from messages import introduction
 
-        introduction()
+        auto_install = auto_install or messages.user_wants_auto_configuration()
+        if auto_install:
+            sys.argv.append("--yes")
+        else:
+            messages.introduction()
 
         from invokeai.frontend.install.invokeai_configure import invokeai_configure
 

installer/lib/messages.py

@@ -7,7 +7,7 @@ import os
 import platform
 from pathlib import Path
 
-from prompt_toolkit import prompt
+from prompt_toolkit import HTML, prompt
 from prompt_toolkit.completion import PathCompleter
 from prompt_toolkit.validation import Validator
 from rich import box, print
@@ -65,17 +65,50 @@ def confirm_install(dest: Path) -> bool:
     if dest.exists():
         print(f":exclamation: Directory {dest} already exists :exclamation:")
         dest_confirmed = Confirm.ask(
-            ":stop_sign: Are you sure you want to (re)install in this location?",
+            ":stop_sign: (re)install in this location?",
             default=False,
         )
     else:
         print(f"InvokeAI will be installed in {dest}")
-        dest_confirmed = not Confirm.ask("Would you like to pick a different location?", default=False)
+        dest_confirmed = Confirm.ask("Use this location?", default=True)
     console.line()
 
     return dest_confirmed
 
 
+def user_wants_auto_configuration() -> bool:
+    """Prompt the user to choose between manual and auto configuration."""
+    console.rule("InvokeAI Configuration Section")
+    console.print(
+        Panel(
+            Group(
+                "\n".join(
+                    [
+                        "Libraries are installed and InvokeAI will now set up its root directory and configuration. Choose between:",
+                        "",
+                        "  * AUTOMATIC configuration:  install reasonable defaults and a minimal set of starter models.",
+                        "  * MANUAL configuration: manually inspect and adjust configuration options and pick from a larger set of starter models.",
+                        "",
+                        "Later you can fine tune your configuration by selecting option [6] 'Change InvokeAI startup options' from the invoke.bat/invoke.sh launcher script.",
+                    ]
+                ),
+            ),
+            box=box.MINIMAL,
+            padding=(1, 1),
+        )
+    )
+    choice = (
+        prompt(
+            HTML("Choose <b>&lt;a&gt;</b>utomatic or <b>&lt;m&gt;</b>anual configuration [a/m] (a): "),
+            validator=Validator.from_callable(
+                lambda n: n == "" or n.startswith(("a", "A", "m", "M")), error_message="Please select 'a' or 'm'"
+            ),
+        )
+        or "a"
+    )
+    return choice.lower().startswith("a")
+
+
 def dest_path(dest=None) -> Path:
     """
     Prompt the user for the destination path and create the path

installer/templates/invoke.bat.in

@@ -17,9 +17,10 @@ echo 6. Change InvokeAI startup options
 echo 7. Re-run the configure script to fix a broken install or to complete a major upgrade
 echo 8. Open the developer console
 echo 9. Update InvokeAI
-echo 10. Command-line help
+echo 10. Run the InvokeAI image database maintenance script
+echo 11. Command-line help
 echo Q - Quit
-set /P choice="Please enter 1-10, Q: [1] "
+set /P choice="Please enter 1-11, Q: [1] "
 if not defined choice set choice=1
 IF /I "%choice%" == "1" (
     echo Starting the InvokeAI browser-based UI..
@@ -58,8 +59,11 @@ IF /I "%choice%" == "1" (
    echo Running invokeai-update...
    python -m invokeai.frontend.install.invokeai_update
 ) ELSE IF /I "%choice%" == "10" (
+   echo Running the db maintenance script...
+   python .venv\Scripts\invokeai-db-maintenance.exe
+) ELSE IF /I "%choice%" == "11" (
     echo Displaying command line help...
-    python .venv\Scripts\invokeai.exe --help %*
+    python .venv\Scripts\invokeai-web.exe --help %*
     pause
     exit /b
 ) ELSE IF /I "%choice%" == "q" (

installer/templates/invoke.sh.in

@@ -46,6 +46,9 @@ if [ "$(uname -s)" == "Darwin" ]; then
     export PYTORCH_ENABLE_MPS_FALLBACK=1
 fi
 
+# Avoid glibc memory fragmentation. See invokeai/backend/model_management/README.md for details.
+export MALLOC_MMAP_THRESHOLD_=1048576
+
 # Primary function for the case statement to determine user input
 do_choice() {
     case $1 in
@@ -97,13 +100,13 @@ do_choice() {
         ;;
     10)
         clear
-        printf "Command-line help\n"
-        invokeai --help
+        printf "Running the db maintenance script\n"
+        invokeai-db-maintenance --root ${INVOKEAI_ROOT}
         ;;
-    "HELP 1")
+    11)
         clear
         printf "Command-line help\n"
-        invokeai --help
+        invokeai-web --help
         ;;
     *)
         clear
@@ -125,7 +128,10 @@ do_dialog() {
         6 "Change InvokeAI startup options"
         7 "Re-run the configure script to fix a broken install or to complete a major upgrade"
         8 "Open the developer console"
-        9 "Update InvokeAI")
+        9 "Update InvokeAI"
+	10 "Run the InvokeAI image database maintenance script"
+	11 "Command-line help"
+    )
 
     choice=$(dialog --clear \
         --backtitle "\Zb\Zu\Z3InvokeAI" \
@@ -157,9 +163,10 @@ do_line_input() {
     printf "7: Re-run the configure script to fix a broken install\n"
     printf "8: Open the developer console\n"
     printf "9: Update InvokeAI\n"
-    printf "10: Command-line help\n"
+    printf "10: Run the InvokeAI image database maintenance script\n"
+    printf "11: Command-line help\n"
     printf "Q: Quit\n\n"
-    read -p "Please enter 1-10, Q: [1] " yn
+    read -p "Please enter 1-11, Q: [1] " yn
     choice=${yn:='1'}
     do_choice $choice
     clear

invokeai/app/api/dependencies.py

@@ -49,7 +49,7 @@ def check_internet() -> bool:
         return False
 
 
-logger = InvokeAILogger.getLogger()
+logger = InvokeAILogger.get_logger()
 
 
 class ApiDependencies:

invokeai/app/api/routers/app_info.py

@@ -7,6 +7,7 @@ from fastapi.routing import APIRouter
 from pydantic import BaseModel, Field
 
 from invokeai.app.invocations.upscale import ESRGAN_MODELS
+from invokeai.app.services.invocation_cache.invocation_cache_common import InvocationCacheStatus
 from invokeai.backend.image_util.invisible_watermark import InvisibleWatermark
 from invokeai.backend.image_util.patchmatch import PatchMatch
 from invokeai.backend.image_util.safety_checker import SafetyChecker
@@ -103,3 +104,43 @@ async def set_log_level(
     """Sets the log verbosity level"""
     ApiDependencies.invoker.services.logger.setLevel(level)
     return LogLevel(ApiDependencies.invoker.services.logger.level)
+
+
+@app_router.delete(
+    "/invocation_cache",
+    operation_id="clear_invocation_cache",
+    responses={200: {"description": "The operation was successful"}},
+)
+async def clear_invocation_cache() -> None:
+    """Clears the invocation cache"""
+    ApiDependencies.invoker.services.invocation_cache.clear()
+
+
+@app_router.put(
+    "/invocation_cache/enable",
+    operation_id="enable_invocation_cache",
+    responses={200: {"description": "The operation was successful"}},
+)
+async def enable_invocation_cache() -> None:
+    """Clears the invocation cache"""
+    ApiDependencies.invoker.services.invocation_cache.enable()
+
+
+@app_router.put(
+    "/invocation_cache/disable",
+    operation_id="disable_invocation_cache",
+    responses={200: {"description": "The operation was successful"}},
+)
+async def disable_invocation_cache() -> None:
+    """Clears the invocation cache"""
+    ApiDependencies.invoker.services.invocation_cache.disable()
+
+
+@app_router.get(
+    "/invocation_cache/status",
+    operation_id="get_invocation_cache_status",
+    responses={200: {"model": InvocationCacheStatus}},
+)
+async def get_invocation_cache_status() -> InvocationCacheStatus:
+    """Clears the invocation cache"""
+    return ApiDependencies.invoker.services.invocation_cache.get_status()

invokeai/app/api/routers/images.py

@@ -322,3 +322,20 @@ async def unstar_images_in_list(
         return ImagesUpdatedFromListResult(updated_image_names=updated_image_names)
     except Exception:
         raise HTTPException(status_code=500, detail="Failed to unstar images")
+
+
+class ImagesDownloaded(BaseModel):
+    response: Optional[str] = Field(
+        description="If defined, the message to display to the user when images begin downloading"
+    )
+
+
+@images_router.post("/download", operation_id="download_images_from_list", response_model=ImagesDownloaded)
+async def download_images_from_list(
+    image_names: list[str] = Body(description="The list of names of images to download", embed=True),
+    board_id: Optional[str] = Body(
+        default=None, description="The board from which image should be downloaded from", embed=True
+    ),
+) -> ImagesDownloaded:
+    # return ImagesDownloaded(response="Your images are downloading")
+    raise HTTPException(status_code=501, detail="Endpoint is not yet implemented")

invokeai/app/api/routers/models.py

@@ -146,7 +146,8 @@ async def update_model(
 async def import_model(
     location: str = Body(description="A model path, repo_id or URL to import"),
     prediction_type: Optional[Literal["v_prediction", "epsilon", "sample"]] = Body(
-        description="Prediction type for SDv2 checkpoint files", default="v_prediction"
+        description="Prediction type for SDv2 checkpoints and rare SDv1 checkpoints",
+        default=None,
     ),
 ) -> ImportModelResponse:
     """Add a model using its local path, repo_id, or remote URL. Model characteristics will be probed and configured automatically"""

invokeai/app/api/routers/session_queue.py

@@ -84,7 +84,7 @@ async def list_queue_items(
     """Gets all queue items (without graphs)"""
 
     return ApiDependencies.invoker.services.session_queue.list_queue_items(
-        queue_id=queue_id, limit=limit, status=status, order_id=cursor, priority=priority
+        queue_id=queue_id, limit=limit, status=status, cursor=cursor, priority=priority
     )
 
 
@@ -225,7 +225,7 @@ async def get_batch_status(
 )
 async def get_queue_item(
     queue_id: str = Path(description="The queue id to perform this operation on"),
-    item_id: str = Path(description="The queue item to get"),
+    item_id: int = Path(description="The queue item to get"),
 ) -> SessionQueueItem:
     """Gets a queue item"""
     return ApiDependencies.invoker.services.session_queue.get_queue_item(item_id)
@@ -240,7 +240,7 @@ async def get_queue_item(
 )
 async def cancel_queue_item(
     queue_id: str = Path(description="The queue id to perform this operation on"),
-    item_id: str = Path(description="The queue item to cancel"),
+    item_id: int = Path(description="The queue item to cancel"),
 ) -> SessionQueueItem:
     """Deletes a queue item"""
 

invokeai/app/api/sockets.py

@@ -3,16 +3,19 @@
 from fastapi import FastAPI
 from fastapi_events.handlers.local import local_handler
 from fastapi_events.typing import Event
-from fastapi_socketio import SocketManager
+from socketio import ASGIApp, AsyncServer
 
 from ..services.events import EventServiceBase
 
 
 class SocketIO:
-    __sio: SocketManager
+    __sio: AsyncServer
+    __app: ASGIApp
 
     def __init__(self, app: FastAPI):
-        self.__sio = SocketManager(app=app)
+        self.__sio = AsyncServer(async_mode="asgi", cors_allowed_origins="*")
+        self.__app = ASGIApp(socketio_server=self.__sio, socketio_path="socket.io")
+        app.mount("/ws", self.__app)
 
         self.__sio.on("subscribe_queue", handler=self._handle_sub_queue)
         self.__sio.on("unsubscribe_queue", handler=self._handle_unsub_queue)

invokeai/app/api_app.py

@@ -8,7 +8,6 @@ app_config.parse_args()
 
 if True:  # hack to make flake8 happy with imports coming after setting up the config
     import asyncio
-    import logging
     import mimetypes
     import socket
     from inspect import signature
@@ -41,7 +40,9 @@ if True:  # hack to make flake8 happy with imports coming after setting up the c
         import invokeai.backend.util.mps_fixes  # noqa: F401 (monkeypatching on import)
 
 
-logger = InvokeAILogger.getLogger(config=app_config)
+app_config = InvokeAIAppConfig.get_config()
+app_config.parse_args()
+logger = InvokeAILogger.get_logger(config=app_config)
 
 # fix for windows mimetypes registry entries being borked
 # see https://github.com/invoke-ai/InvokeAI/discussions/3684#discussioncomment-6391352
@@ -223,7 +224,7 @@ def invoke_api():
                 exc_info=e,
             )
         else:
-            jurigged.watch(logger=InvokeAILogger.getLogger(name="jurigged").info)
+            jurigged.watch(logger=InvokeAILogger.get_logger(name="jurigged").info)
 
     port = find_port(app_config.port)
     if port != app_config.port:
@@ -242,7 +243,7 @@ def invoke_api():
 
     # replace uvicorn's loggers with InvokeAI's for consistent appearance
     for logname in ["uvicorn.access", "uvicorn"]:
-        log = logging.getLogger(logname)
+        log = InvokeAILogger.get_logger(logname)
         log.handlers.clear()
         for ch in logger.handlers:
             log.addHandler(ch)

invokeai/app/cli_app.py

@@ -7,8 +7,6 @@ from .services.config import InvokeAIAppConfig
 # parse_args() must be called before any other imports. if it is not called first, consumers of the config
 # which are imported/used before parse_args() is called will get the default config values instead of the
 # values from the command line or config file.
-config = InvokeAIAppConfig.get_config()
-config.parse_args()
 
 if True:  # hack to make flake8 happy with imports coming after setting up the config
     import argparse
@@ -61,8 +59,9 @@ if True:  # hack to make flake8 happy with imports coming after setting up the c
     if torch.backends.mps.is_available():
         import invokeai.backend.util.mps_fixes  # noqa: F401 (monkeypatching on import)
 
-
-logger = InvokeAILogger().getLogger(config=config)
+config = InvokeAIAppConfig.get_config()
+config.parse_args()
+logger = InvokeAILogger().get_logger(config=config)
 
 
 class CliCommand(BaseModel):

invokeai/app/invocations/baseinvocation.py

@@ -67,6 +67,8 @@ class FieldDescriptions:
     width = "Width of output (px)"
     height = "Height of output (px)"
     control = "ControlNet(s) to apply"
+    ip_adapter = "IP-Adapter to apply"
+    t2i_adapter = "T2I-Adapter(s) to apply"
     denoised_latents = "Denoised latents tensor"
     latents = "Latents tensor"
     strength = "Strength of denoising (proportional to steps)"
@@ -87,6 +89,12 @@ class FieldDescriptions:
     num_1 = "The first number"
     num_2 = "The second number"
     mask = "The mask to use for the operation"
+    board = "The board to save the image to"
+    image = "The image to process"
+    tile_size = "Tile size"
+    inclusive_low = "The inclusive low value"
+    exclusive_high = "The exclusive high value"
+    decimal_places = "The number of decimal places to round to"
 
 
 class Input(str, Enum):
@@ -155,6 +163,7 @@ class UIType(str, Enum):
     VaeModel = "VaeModelField"
     LoRAModel = "LoRAModelField"
     ControlNetModel = "ControlNetModelField"
+    IPAdapterModel = "IPAdapterModelField"
     UNet = "UNetField"
     Vae = "VaeField"
     CLIP = "ClipField"
@@ -171,6 +180,7 @@ class UIType(str, Enum):
     WorkflowField = "WorkflowField"
     IsIntermediate = "IsIntermediate"
     MetadataField = "MetadataField"
+    BoardField = "BoardField"
     # endregion
 
 
@@ -422,13 +432,22 @@ class InvocationContext:
     services: InvocationServices
     graph_execution_state_id: str
     queue_id: str
-    queue_item_id: str
+    queue_item_id: int
+    queue_batch_id: str
 
-    def __init__(self, services: InvocationServices, queue_id: str, queue_item_id: str, graph_execution_state_id: str):
+    def __init__(
+        self,
+        services: InvocationServices,
+        queue_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
+        graph_execution_state_id: str,
+    ):
         self.services = services
         self.graph_execution_state_id = graph_execution_state_id
         self.queue_id = queue_id
         self.queue_item_id = queue_item_id
+        self.queue_batch_id = queue_batch_id
 
 
 class BaseInvocationOutput(BaseModel):
@@ -589,7 +608,7 @@ class BaseInvocation(ABC, BaseModel):
             if cached_value is None:
                 context.services.logger.debug(f'Invocation cache miss for type "{self.get_type()}": {self.id}')
                 output = self.invoke(context)
-                context.services.invocation_cache.save(output)
+                context.services.invocation_cache.save(key, output)
                 return output
             else:
                 context.services.logger.debug(f'Invocation cache hit for type "{self.get_type()}": {self.id}')
@@ -645,6 +664,8 @@ def invocation(
     :param Optional[str] title: Adds a title to the invocation. Use if the auto-generated title isn't quite right. Defaults to None.
     :param Optional[list[str]] tags: Adds tags to the invocation. Invocations may be searched for by their tags. Defaults to None.
     :param Optional[str] category: Adds a category to the invocation. Used to group the invocations in the UI. Defaults to None.
+    :param Optional[str] version: Adds a version to the invocation. Must be a valid semver string. Defaults to None.
+    :param Optional[bool] use_cache: Whether or not to use the invocation cache. Defaults to True. The user may override this in the workflow editor.
     """
 
     def wrapper(cls: Type[GenericBaseInvocation]) -> Type[GenericBaseInvocation]:

invokeai/app/invocations/compel.py

@@ -7,14 +7,14 @@ from compel import Compel, ReturnedEmbeddingsType
 from compel.prompt_parser import Blend, Conjunction, CrossAttentionControlSubstitute, FlattenedPrompt, Fragment
 
 from invokeai.app.invocations.primitives import ConditioningField, ConditioningOutput
-from invokeai.backend.stable_diffusion.diffusion.shared_invokeai_diffusion import (
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
     BasicConditioningInfo,
+    ExtraConditioningInfo,
     SDXLConditioningInfo,
 )
 
 from ...backend.model_management.lora import ModelPatcher
 from ...backend.model_management.models import ModelNotFoundException, ModelType
-from ...backend.stable_diffusion.diffusion import InvokeAIDiffuserComponent
 from ...backend.util.devices import torch_dtype
 from .baseinvocation import (
     BaseInvocation,
@@ -99,14 +99,15 @@ class CompelInvocation(BaseInvocation):
                 # print(traceback.format_exc())
                 print(f'Warn: trigger: "{trigger}" not found')
 
-        with ModelPatcher.apply_lora_text_encoder(
-            text_encoder_info.context.model, _lora_loader()
-        ), ModelPatcher.apply_ti(tokenizer_info.context.model, text_encoder_info.context.model, ti_list) as (
-            tokenizer,
-            ti_manager,
-        ), ModelPatcher.apply_clip_skip(
-            text_encoder_info.context.model, self.clip.skipped_layers
-        ), text_encoder_info as text_encoder:
+        with (
+            ModelPatcher.apply_lora_text_encoder(text_encoder_info.context.model, _lora_loader()),
+            ModelPatcher.apply_ti(tokenizer_info.context.model, text_encoder_info.context.model, ti_list) as (
+                tokenizer,
+                ti_manager,
+            ),
+            ModelPatcher.apply_clip_skip(text_encoder_info.context.model, self.clip.skipped_layers),
+            text_encoder_info as text_encoder,
+        ):
             compel = Compel(
                 tokenizer=tokenizer,
                 text_encoder=text_encoder,
@@ -122,7 +123,7 @@ class CompelInvocation(BaseInvocation):
 
             c, options = compel.build_conditioning_tensor_for_conjunction(conjunction)
 
-            ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(
+            ec = ExtraConditioningInfo(
                 tokens_count_including_eos_bos=get_max_token_count(tokenizer, conjunction),
                 cross_attention_control_args=options.get("cross_attention_control", None),
             )
@@ -213,14 +214,15 @@ class SDXLPromptInvocationBase:
                 # print(traceback.format_exc())
                 print(f'Warn: trigger: "{trigger}" not found')
 
-        with ModelPatcher.apply_lora(
-            text_encoder_info.context.model, _lora_loader(), lora_prefix
-        ), ModelPatcher.apply_ti(tokenizer_info.context.model, text_encoder_info.context.model, ti_list) as (
-            tokenizer,
-            ti_manager,
-        ), ModelPatcher.apply_clip_skip(
-            text_encoder_info.context.model, clip_field.skipped_layers
-        ), text_encoder_info as text_encoder:
+        with (
+            ModelPatcher.apply_lora(text_encoder_info.context.model, _lora_loader(), lora_prefix),
+            ModelPatcher.apply_ti(tokenizer_info.context.model, text_encoder_info.context.model, ti_list) as (
+                tokenizer,
+                ti_manager,
+            ),
+            ModelPatcher.apply_clip_skip(text_encoder_info.context.model, clip_field.skipped_layers),
+            text_encoder_info as text_encoder,
+        ):
             compel = Compel(
                 tokenizer=tokenizer,
                 text_encoder=text_encoder,
@@ -244,7 +246,7 @@ class SDXLPromptInvocationBase:
             else:
                 c_pooled = None
 
-            ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(
+            ec = ExtraConditioningInfo(
                 tokens_count_including_eos_bos=get_max_token_count(tokenizer, conjunction),
                 cross_attention_control_args=options.get("cross_attention_control", None),
             )
@@ -436,9 +438,11 @@ def get_tokens_for_prompt_object(tokenizer, parsed_prompt: FlattenedPrompt, trun
         raise ValueError("Blend is not supported here - you need to get tokens for each of its .children")
 
     text_fragments = [
-        x.text
-        if type(x) is Fragment
-        else (" ".join([f.text for f in x.original]) if type(x) is CrossAttentionControlSubstitute else str(x))
+        (
+            x.text
+            if type(x) is Fragment
+            else (" ".join([f.text for f in x.original]) if type(x) is CrossAttentionControlSubstitute else str(x))
+        )
         for x in parsed_prompt.children
     ]
     text = " ".join(text_fragments)

invokeai/app/invocations/controlnet_image_processors.py

@@ -38,7 +38,6 @@ from .baseinvocation import (
     InputField,
     InvocationContext,
     OutputField,
-    UIType,
     invocation,
     invocation_output,
 )
@@ -100,7 +99,7 @@ class ControlNetInvocation(BaseInvocation):
     image: ImageField = InputField(description="The control image")
     control_model: ControlNetModelField = InputField(description=FieldDescriptions.controlnet_model, input=Input.Direct)
     control_weight: Union[float, List[float]] = InputField(
-        default=1.0, description="The weight given to the ControlNet", ui_type=UIType.Float
+        default=1.0, description="The weight given to the ControlNet"
     )
     begin_step_percent: float = InputField(
         default=0, ge=-1, le=2, description="When the ControlNet is first applied (% of total steps)"
@@ -560,3 +559,33 @@ class SamDetectorReproducibleColors(SamDetector):
             img[:, :] = ann_color
             final_img.paste(Image.fromarray(img, mode="RGB"), (0, 0), Image.fromarray(np.uint8(m * 255)))
         return np.array(final_img, dtype=np.uint8)
+
+
+@invocation(
+    "color_map_image_processor",
+    title="Color Map Processor",
+    tags=["controlnet"],
+    category="controlnet",
+    version="1.0.0",
+)
+class ColorMapImageProcessorInvocation(ImageProcessorInvocation):
+    """Generates a color map from the provided image"""
+
+    color_map_tile_size: int = InputField(default=64, ge=0, description=FieldDescriptions.tile_size)
+
+    def run_processor(self, image: Image.Image):
+        image = image.convert("RGB")
+        image = np.array(image, dtype=np.uint8)
+        height, width = image.shape[:2]
+
+        width_tile_size = min(self.color_map_tile_size, width)
+        height_tile_size = min(self.color_map_tile_size, height)
+
+        color_map = cv2.resize(
+            image,
+            (width // width_tile_size, height // height_tile_size),
+            interpolation=cv2.INTER_CUBIC,
+        )
+        color_map = cv2.resize(color_map, (width, height), interpolation=cv2.INTER_NEAREST)
+        color_map = Image.fromarray(color_map)
+        return color_map

invokeai/app/invocations/facetools.py

@@ -0,0 +1,692 @@
+import math
+import re
+from pathlib import Path
+from typing import Optional, TypedDict
+
+import cv2
+import numpy as np
+from mediapipe.python.solutions.face_mesh import FaceMesh  # type: ignore[import]
+from PIL import Image, ImageDraw, ImageFilter, ImageFont, ImageOps
+from PIL.Image import Image as ImageType
+from pydantic import validator
+
+import invokeai.assets.fonts as font_assets
+from invokeai.app.invocations.baseinvocation import (
+    BaseInvocation,
+    InputField,
+    InvocationContext,
+    OutputField,
+    invocation,
+    invocation_output,
+)
+from invokeai.app.invocations.primitives import ImageField, ImageOutput
+from invokeai.app.models.image import ImageCategory, ResourceOrigin
+
+
+@invocation_output("face_mask_output")
+class FaceMaskOutput(ImageOutput):
+    """Base class for FaceMask output"""
+
+    mask: ImageField = OutputField(description="The output mask")
+
+
+@invocation_output("face_off_output")
+class FaceOffOutput(ImageOutput):
+    """Base class for FaceOff Output"""
+
+    mask: ImageField = OutputField(description="The output mask")
+    x: int = OutputField(description="The x coordinate of the bounding box's left side")
+    y: int = OutputField(description="The y coordinate of the bounding box's top side")
+
+
+class FaceResultData(TypedDict):
+    image: ImageType
+    mask: ImageType
+    x_center: float
+    y_center: float
+    mesh_width: int
+    mesh_height: int
+
+
+class FaceResultDataWithId(FaceResultData):
+    face_id: int
+
+
+class ExtractFaceData(TypedDict):
+    bounded_image: ImageType
+    bounded_mask: ImageType
+    x_min: int
+    y_min: int
+    x_max: int
+    y_max: int
+
+
+class FaceMaskResult(TypedDict):
+    image: ImageType
+    mask: ImageType
+
+
+def create_white_image(w: int, h: int) -> ImageType:
+    return Image.new("L", (w, h), color=255)
+
+
+def create_black_image(w: int, h: int) -> ImageType:
+    return Image.new("L", (w, h), color=0)
+
+
+FONT_SIZE = 32
+FONT_STROKE_WIDTH = 4
+
+
+def prepare_faces_list(
+    face_result_list: list[FaceResultData],
+) -> list[FaceResultDataWithId]:
+    """Deduplicates a list of faces, adding IDs to them."""
+    deduped_faces: list[FaceResultData] = []
+
+    if len(face_result_list) == 0:
+        return list()
+
+    for candidate in face_result_list:
+        should_add = True
+        candidate_x_center = candidate["x_center"]
+        candidate_y_center = candidate["y_center"]
+        for face in deduped_faces:
+            face_center_x = face["x_center"]
+            face_center_y = face["y_center"]
+            face_radius_w = face["mesh_width"] / 2
+            face_radius_h = face["mesh_height"] / 2
+            # Determine if the center of the candidate_face is inside the ellipse of the added face
+            # p < 1 -> Inside
+            # p = 1 -> Exactly on the ellipse
+            # p > 1 -> Outside
+            p = (math.pow((candidate_x_center - face_center_x), 2) / math.pow(face_radius_w, 2)) + (
+                math.pow((candidate_y_center - face_center_y), 2) / math.pow(face_radius_h, 2)
+            )
+
+            if p < 1:  # Inside of the already-added face's radius
+                should_add = False
+                break
+
+        if should_add is True:
+            deduped_faces.append(candidate)
+
+    sorted_faces = sorted(deduped_faces, key=lambda x: x["y_center"])
+    sorted_faces = sorted(sorted_faces, key=lambda x: x["x_center"])
+
+    # add face_id for reference
+    sorted_faces_with_ids: list[FaceResultDataWithId] = []
+    face_id_counter = 0
+    for face in sorted_faces:
+        sorted_faces_with_ids.append(
+            FaceResultDataWithId(
+                **face,
+                face_id=face_id_counter,
+            )
+        )
+        face_id_counter += 1
+
+    return sorted_faces_with_ids
+
+
+def generate_face_box_mask(
+    context: InvocationContext,
+    minimum_confidence: float,
+    x_offset: float,
+    y_offset: float,
+    pil_image: ImageType,
+    chunk_x_offset: int = 0,
+    chunk_y_offset: int = 0,
+    draw_mesh: bool = True,
+    check_bounds: bool = True,
+) -> list[FaceResultData]:
+    result = []
+    mask_pil = None
+
+    # Convert the PIL image to a NumPy array.
+    np_image = np.array(pil_image, dtype=np.uint8)
+
+    # Check if the input image has four channels (RGBA).
+    if np_image.shape[2] == 4:
+        # Convert RGBA to RGB by removing the alpha channel.
+        np_image = np_image[:, :, :3]
+
+    # Create a FaceMesh object for face landmark detection and mesh generation.
+    face_mesh = FaceMesh(
+        max_num_faces=999,
+        min_detection_confidence=minimum_confidence,
+        min_tracking_confidence=minimum_confidence,
+    )
+
+    # Detect the face landmarks and mesh in the input image.
+    results = face_mesh.process(np_image)
+
+    # Check if any face is detected.
+    if results.multi_face_landmarks:  # type: ignore # this are via protobuf and not typed
+        # Search for the face_id in the detected faces.
+        for face_id, face_landmarks in enumerate(results.multi_face_landmarks):  # type: ignore #this are via protobuf and not typed
+            # Get the bounding box of the face mesh.
+            x_coordinates = [landmark.x for landmark in face_landmarks.landmark]
+            y_coordinates = [landmark.y for landmark in face_landmarks.landmark]
+            x_min, x_max = min(x_coordinates), max(x_coordinates)
+            y_min, y_max = min(y_coordinates), max(y_coordinates)
+
+            # Calculate the width and height of the face mesh.
+            mesh_width = int((x_max - x_min) * np_image.shape[1])
+            mesh_height = int((y_max - y_min) * np_image.shape[0])
+
+            # Get the center of the face.
+            x_center = np.mean([landmark.x * np_image.shape[1] for landmark in face_landmarks.landmark])
+            y_center = np.mean([landmark.y * np_image.shape[0] for landmark in face_landmarks.landmark])
+
+            face_landmark_points = np.array(
+                [
+                    [landmark.x * np_image.shape[1], landmark.y * np_image.shape[0]]
+                    for landmark in face_landmarks.landmark
+                ]
+            )
+
+            # Apply the scaling offsets to the face landmark points with a multiplier.
+            scale_multiplier = 0.2
+            x_center = np.mean(face_landmark_points[:, 0])
+            y_center = np.mean(face_landmark_points[:, 1])
+
+            if draw_mesh:
+                x_scaled = face_landmark_points[:, 0] + scale_multiplier * x_offset * (
+                    face_landmark_points[:, 0] - x_center
+                )
+                y_scaled = face_landmark_points[:, 1] + scale_multiplier * y_offset * (
+                    face_landmark_points[:, 1] - y_center
+                )
+
+                convex_hull = cv2.convexHull(np.column_stack((x_scaled, y_scaled)).astype(np.int32))
+
+                # Generate a binary face mask using the face mesh.
+                mask_image = np.ones(np_image.shape[:2], dtype=np.uint8) * 255
+                cv2.fillConvexPoly(mask_image, convex_hull, 0)
+
+                # Convert the binary mask image to a PIL Image.
+                init_mask_pil = Image.fromarray(mask_image, mode="L")
+                w, h = init_mask_pil.size
+                mask_pil = create_white_image(w + chunk_x_offset, h + chunk_y_offset)
+                mask_pil.paste(init_mask_pil, (chunk_x_offset, chunk_y_offset))
+
+            left_side = x_center - mesh_width
+            right_side = x_center + mesh_width
+            top_side = y_center - mesh_height
+            bottom_side = y_center + mesh_height
+            im_width, im_height = pil_image.size
+            over_w = im_width * 0.1
+            over_h = im_height * 0.1
+            if not check_bounds or (
+                (left_side >= -over_w)
+                and (right_side < im_width + over_w)
+                and (top_side >= -over_h)
+                and (bottom_side < im_height + over_h)
+            ):
+                x_center = float(x_center)
+                y_center = float(y_center)
+                face = FaceResultData(
+                    image=pil_image,
+                    mask=mask_pil or create_white_image(*pil_image.size),
+                    x_center=x_center + chunk_x_offset,
+                    y_center=y_center + chunk_y_offset,
+                    mesh_width=mesh_width,
+                    mesh_height=mesh_height,
+                )
+
+                result.append(face)
+            else:
+                context.services.logger.info("FaceTools --> Face out of bounds, ignoring.")
+
+    return result
+
+
+def extract_face(
+    context: InvocationContext,
+    image: ImageType,
+    face: FaceResultData,
+    padding: int,
+) -> ExtractFaceData:
+    mask = face["mask"]
+    center_x = face["x_center"]
+    center_y = face["y_center"]
+    mesh_width = face["mesh_width"]
+    mesh_height = face["mesh_height"]
+
+    # Determine the minimum size of the square crop
+    min_size = min(mask.width, mask.height)
+
+    # Calculate the crop boundaries for the output image and mask.
+    mesh_width += 128 + padding  # add pixels to account for mask variance
+    mesh_height += 128 + padding  # add pixels to account for mask variance
+    crop_size = min(
+        max(mesh_width, mesh_height, 128), min_size
+    )  # Choose the smaller of the two (given value or face mask size)
+    if crop_size > 128:
+        crop_size = (crop_size + 7) // 8 * 8  # Ensure crop side is multiple of 8
+
+    # Calculate the actual crop boundaries within the bounds of the original image.
+    x_min = int(center_x - crop_size / 2)
+    y_min = int(center_y - crop_size / 2)
+    x_max = int(center_x + crop_size / 2)
+    y_max = int(center_y + crop_size / 2)
+
+    # Adjust the crop boundaries to stay within the original image's dimensions
+    if x_min < 0:
+        context.services.logger.warning("FaceTools --> -X-axis padding reached image edge.")
+        x_max -= x_min
+        x_min = 0
+    elif x_max > mask.width:
+        context.services.logger.warning("FaceTools --> +X-axis padding reached image edge.")
+        x_min -= x_max - mask.width
+        x_max = mask.width
+
+    if y_min < 0:
+        context.services.logger.warning("FaceTools --> +Y-axis padding reached image edge.")
+        y_max -= y_min
+        y_min = 0
+    elif y_max > mask.height:
+        context.services.logger.warning("FaceTools --> -Y-axis padding reached image edge.")
+        y_min -= y_max - mask.height
+        y_max = mask.height
+
+    # Ensure the crop is square and adjust the boundaries if needed
+    if x_max - x_min != crop_size:
+        context.services.logger.warning("FaceTools --> Limiting x-axis padding to constrain bounding box to a square.")
+        diff = crop_size - (x_max - x_min)
+        x_min -= diff // 2
+        x_max += diff - diff // 2
+
+    if y_max - y_min != crop_size:
+        context.services.logger.warning("FaceTools --> Limiting y-axis padding to constrain bounding box to a square.")
+        diff = crop_size - (y_max - y_min)
+        y_min -= diff // 2
+        y_max += diff - diff // 2
+
+    context.services.logger.info(f"FaceTools --> Calculated bounding box (8 multiple): {crop_size}")
+
+    # Crop the output image to the specified size with the center of the face mesh as the center.
+    mask = mask.crop((x_min, y_min, x_max, y_max))
+    bounded_image = image.crop((x_min, y_min, x_max, y_max))
+
+    # blur mask edge by small radius
+    mask = mask.filter(ImageFilter.GaussianBlur(radius=2))
+
+    return ExtractFaceData(
+        bounded_image=bounded_image,
+        bounded_mask=mask,
+        x_min=x_min,
+        y_min=y_min,
+        x_max=x_max,
+        y_max=y_max,
+    )
+
+
+def get_faces_list(
+    context: InvocationContext,
+    image: ImageType,
+    should_chunk: bool,
+    minimum_confidence: float,
+    x_offset: float,
+    y_offset: float,
+    draw_mesh: bool = True,
+) -> list[FaceResultDataWithId]:
+    result = []
+
+    # Generate the face box mask and get the center of the face.
+    if not should_chunk:
+        context.services.logger.info("FaceTools --> Attempting full image face detection.")
+        result = generate_face_box_mask(
+            context=context,
+            minimum_confidence=minimum_confidence,
+            x_offset=x_offset,
+            y_offset=y_offset,
+            pil_image=image,
+            chunk_x_offset=0,
+            chunk_y_offset=0,
+            draw_mesh=draw_mesh,
+            check_bounds=False,
+        )
+    if should_chunk or len(result) == 0:
+        context.services.logger.info("FaceTools --> Chunking image (chunk toggled on, or no face found in full image).")
+        width, height = image.size
+        image_chunks = []
+        x_offsets = []
+        y_offsets = []
+        result = []
+
+        # If width == height, there's nothing more we can do... otherwise...
+        if width > height:
+            # Landscape - slice the image horizontally
+            fx = 0.0
+            steps = int(width * 2 / height)
+            while fx <= (width - height):
+                x = int(fx)
+                image_chunks.append(image.crop((x, 0, x + height - 1, height - 1)))
+                x_offsets.append(x)
+                y_offsets.append(0)
+                fx += (width - height) / steps
+                context.services.logger.info(f"FaceTools --> Chunk starting at x = {x}")
+        elif height > width:
+            # Portrait - slice the image vertically
+            fy = 0.0
+            steps = int(height * 2 / width)
+            while fy <= (height - width):
+                y = int(fy)
+                image_chunks.append(image.crop((0, y, width - 1, y + width - 1)))
+                x_offsets.append(0)
+                y_offsets.append(y)
+                fy += (height - width) / steps
+                context.services.logger.info(f"FaceTools --> Chunk starting at y = {y}")
+
+        for idx in range(len(image_chunks)):
+            context.services.logger.info(f"FaceTools --> Evaluating faces in chunk {idx}")
+            result = result + generate_face_box_mask(
+                context=context,
+                minimum_confidence=minimum_confidence,
+                x_offset=x_offset,
+                y_offset=y_offset,
+                pil_image=image_chunks[idx],
+                chunk_x_offset=x_offsets[idx],
+                chunk_y_offset=y_offsets[idx],
+                draw_mesh=draw_mesh,
+            )
+
+        if len(result) == 0:
+            # Give up
+            context.services.logger.warning(
+                "FaceTools --> No face detected in chunked input image. Passing through original image."
+            )
+
+    all_faces = prepare_faces_list(result)
+
+    return all_faces
+
+
+@invocation("face_off", title="FaceOff", tags=["image", "faceoff", "face", "mask"], category="image", version="1.0.1")
+class FaceOffInvocation(BaseInvocation):
+    """Bound, extract, and mask a face from an image using MediaPipe detection"""
+
+    image: ImageField = InputField(description="Image for face detection")
+    face_id: int = InputField(
+        default=0,
+        ge=0,
+        description="The face ID to process, numbered from 0. Multiple faces not supported. Find a face's ID with FaceIdentifier node.",
+    )
+    minimum_confidence: float = InputField(
+        default=0.5, description="Minimum confidence for face detection (lower if detection is failing)"
+    )
+    x_offset: float = InputField(default=0.0, description="X-axis offset of the mask")
+    y_offset: float = InputField(default=0.0, description="Y-axis offset of the mask")
+    padding: int = InputField(default=0, description="All-axis padding around the mask in pixels")
+    chunk: bool = InputField(
+        default=False,
+        description="Whether to bypass full image face detection and default to image chunking. Chunking will occur if no faces are found in the full image.",
+    )
+
+    def faceoff(self, context: InvocationContext, image: ImageType) -> Optional[ExtractFaceData]:
+        all_faces = get_faces_list(
+            context=context,
+            image=image,
+            should_chunk=self.chunk,
+            minimum_confidence=self.minimum_confidence,
+            x_offset=self.x_offset,
+            y_offset=self.y_offset,
+            draw_mesh=True,
+        )
+
+        if len(all_faces) == 0:
+            context.services.logger.warning("FaceOff --> No faces detected. Passing through original image.")
+            return None
+
+        if self.face_id > len(all_faces) - 1:
+            context.services.logger.warning(
+                f"FaceOff --> Face ID {self.face_id} is outside of the number of faces detected ({len(all_faces)}). Passing through original image."
+            )
+            return None
+
+        face_data = extract_face(context=context, image=image, face=all_faces[self.face_id], padding=self.padding)
+        # Convert the input image to RGBA mode to ensure it has an alpha channel.
+        face_data["bounded_image"] = face_data["bounded_image"].convert("RGBA")
+
+        return face_data
+
+    def invoke(self, context: InvocationContext) -> FaceOffOutput:
+        image = context.services.images.get_pil_image(self.image.image_name)
+        result = self.faceoff(context=context, image=image)
+
+        if result is None:
+            result_image = image
+            result_mask = create_white_image(*image.size)
+            x = 0
+            y = 0
+        else:
+            result_image = result["bounded_image"]
+            result_mask = result["bounded_mask"]
+            x = result["x_min"]
+            y = result["y_min"]
+
+        image_dto = context.services.images.create(
+            image=result_image,
+            image_origin=ResourceOrigin.INTERNAL,
+            image_category=ImageCategory.GENERAL,
+            node_id=self.id,
+            session_id=context.graph_execution_state_id,
+            is_intermediate=self.is_intermediate,
+            workflow=self.workflow,
+        )
+
+        mask_dto = context.services.images.create(
+            image=result_mask,
+            image_origin=ResourceOrigin.INTERNAL,
+            image_category=ImageCategory.MASK,
+            node_id=self.id,
+            session_id=context.graph_execution_state_id,
+            is_intermediate=self.is_intermediate,
+        )
+
+        output = FaceOffOutput(
+            image=ImageField(image_name=image_dto.image_name),
+            width=image_dto.width,
+            height=image_dto.height,
+            mask=ImageField(image_name=mask_dto.image_name),
+            x=x,
+            y=y,
+        )
+
+        return output
+
+
+@invocation("face_mask_detection", title="FaceMask", tags=["image", "face", "mask"], category="image", version="1.0.1")
+class FaceMaskInvocation(BaseInvocation):
+    """Face mask creation using mediapipe face detection"""
+
+    image: ImageField = InputField(description="Image to face detect")
+    face_ids: str = InputField(
+        default="",
+        description="Comma-separated list of face ids to mask eg '0,2,7'. Numbered from 0. Leave empty to mask all. Find face IDs with FaceIdentifier node.",
+    )
+    minimum_confidence: float = InputField(
+        default=0.5, description="Minimum confidence for face detection (lower if detection is failing)"
+    )
+    x_offset: float = InputField(default=0.0, description="Offset for the X-axis of the face mask")
+    y_offset: float = InputField(default=0.0, description="Offset for the Y-axis of the face mask")
+    chunk: bool = InputField(
+        default=False,
+        description="Whether to bypass full image face detection and default to image chunking. Chunking will occur if no faces are found in the full image.",
+    )
+    invert_mask: bool = InputField(default=False, description="Toggle to invert the mask")
+
+    @validator("face_ids")
+    def validate_comma_separated_ints(cls, v) -> str:
+        comma_separated_ints_regex = re.compile(r"^\d*(,\d+)*$")
+        if comma_separated_ints_regex.match(v) is None:
+            raise ValueError('Face IDs must be a comma-separated list of integers (e.g. "1,2,3")')
+        return v
+
+    def facemask(self, context: InvocationContext, image: ImageType) -> FaceMaskResult:
+        all_faces = get_faces_list(
+            context=context,
+            image=image,
+            should_chunk=self.chunk,
+            minimum_confidence=self.minimum_confidence,
+            x_offset=self.x_offset,
+            y_offset=self.y_offset,
+            draw_mesh=True,
+        )
+
+        mask_pil = create_white_image(*image.size)
+
+        id_range = list(range(0, len(all_faces)))
+        ids_to_extract = id_range
+        if self.face_ids != "":
+            parsed_face_ids = [int(id) for id in self.face_ids.split(",")]
+            # get requested face_ids that are in range
+            intersected_face_ids = set(parsed_face_ids) & set(id_range)
+
+            if len(intersected_face_ids) == 0:
+                id_range_str = ",".join([str(id) for id in id_range])
+                context.services.logger.warning(
+                    f"Face IDs must be in range of detected faces - requested {self.face_ids}, detected {id_range_str}. Passing through original image."
+                )
+                return FaceMaskResult(
+                    image=image,  # original image
+                    mask=mask_pil,  # white mask
+                )
+
+            ids_to_extract = list(intersected_face_ids)
+
+        for face_id in ids_to_extract:
+            face_data = extract_face(context=context, image=image, face=all_faces[face_id], padding=0)
+            face_mask_pil = face_data["bounded_mask"]
+            x_min = face_data["x_min"]
+            y_min = face_data["y_min"]
+            x_max = face_data["x_max"]
+            y_max = face_data["y_max"]
+
+            mask_pil.paste(
+                create_black_image(x_max - x_min, y_max - y_min),
+                box=(x_min, y_min),
+                mask=ImageOps.invert(face_mask_pil),
+            )
+
+        if self.invert_mask:
+            mask_pil = ImageOps.invert(mask_pil)
+
+        # Create an RGBA image with transparency
+        image = image.convert("RGBA")
+
+        return FaceMaskResult(
+            image=image,
+            mask=mask_pil,
+        )
+
+    def invoke(self, context: InvocationContext) -> FaceMaskOutput:
+        image = context.services.images.get_pil_image(self.image.image_name)
+        result = self.facemask(context=context, image=image)
+
+        image_dto = context.services.images.create(
+            image=result["image"],
+            image_origin=ResourceOrigin.INTERNAL,
+            image_category=ImageCategory.GENERAL,
+            node_id=self.id,
+            session_id=context.graph_execution_state_id,
+            is_intermediate=self.is_intermediate,
+            workflow=self.workflow,
+        )
+
+        mask_dto = context.services.images.create(
+            image=result["mask"],
+            image_origin=ResourceOrigin.INTERNAL,
+            image_category=ImageCategory.MASK,
+            node_id=self.id,
+            session_id=context.graph_execution_state_id,
+            is_intermediate=self.is_intermediate,
+        )
+
+        output = FaceMaskOutput(
+            image=ImageField(image_name=image_dto.image_name),
+            width=image_dto.width,
+            height=image_dto.height,
+            mask=ImageField(image_name=mask_dto.image_name),
+        )
+
+        return output
+
+
+@invocation(
+    "face_identifier", title="FaceIdentifier", tags=["image", "face", "identifier"], category="image", version="1.0.1"
+)
+class FaceIdentifierInvocation(BaseInvocation):
+    """Outputs an image with detected face IDs printed on each face. For use with other FaceTools."""
+
+    image: ImageField = InputField(description="Image to face detect")
+    minimum_confidence: float = InputField(
+        default=0.5, description="Minimum confidence for face detection (lower if detection is failing)"
+    )
+    chunk: bool = InputField(
+        default=False,
+        description="Whether to bypass full image face detection and default to image chunking. Chunking will occur if no faces are found in the full image.",
+    )
+
+    def faceidentifier(self, context: InvocationContext, image: ImageType) -> ImageType:
+        image = image.copy()
+
+        all_faces = get_faces_list(
+            context=context,
+            image=image,
+            should_chunk=self.chunk,
+            minimum_confidence=self.minimum_confidence,
+            x_offset=0,
+            y_offset=0,
+            draw_mesh=False,
+        )
+
+        # Note - font may be found either in the repo if running an editable install, or in the venv if running a package install
+        font_path = [x for x in [Path(y, "inter/Inter-Regular.ttf") for y in font_assets.__path__] if x.exists()]
+        font = ImageFont.truetype(font_path[0].as_posix(), FONT_SIZE)
+
+        # Paste face IDs on the output image
+        draw = ImageDraw.Draw(image)
+        for face in all_faces:
+            x_coord = face["x_center"]
+            y_coord = face["y_center"]
+            text = str(face["face_id"])
+            # get bbox of the text so we can center the id on the face
+            _, _, bbox_w, bbox_h = draw.textbbox(xy=(0, 0), text=text, font=font, stroke_width=FONT_STROKE_WIDTH)
+            x = x_coord - bbox_w / 2
+            y = y_coord - bbox_h / 2
+            draw.text(
+                xy=(x, y),
+                text=str(text),
+                fill=(255, 255, 255, 255),
+                font=font,
+                stroke_width=FONT_STROKE_WIDTH,
+                stroke_fill=(0, 0, 0, 255),
+            )
+
+        # Create an RGBA image with transparency
+        image = image.convert("RGBA")
+
+        return image
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        image = context.services.images.get_pil_image(self.image.image_name)
+        result_image = self.faceidentifier(context=context, image=image)
+
+        image_dto = context.services.images.create(
+            image=result_image,
+            image_origin=ResourceOrigin.INTERNAL,
+            image_category=ImageCategory.GENERAL,
+            node_id=self.id,
+            session_id=context.graph_execution_state_id,
+            is_intermediate=self.is_intermediate,
+            workflow=self.workflow,
+        )
+
+        return ImageOutput(
+            image=ImageField(image_name=image_dto.image_name),
+            width=image_dto.width,
+            height=image_dto.height,
+        )

invokeai/app/invocations/image.py

@@ -8,12 +8,12 @@ import numpy
 from PIL import Image, ImageChops, ImageFilter, ImageOps
 
 from invokeai.app.invocations.metadata import CoreMetadata
-from invokeai.app.invocations.primitives import ColorField, ImageField, ImageOutput
+from invokeai.app.invocations.primitives import BoardField, ColorField, ImageField, ImageOutput
 from invokeai.backend.image_util.invisible_watermark import InvisibleWatermark
 from invokeai.backend.image_util.safety_checker import SafetyChecker
 
 from ..models.image import ImageCategory, ResourceOrigin
-from .baseinvocation import BaseInvocation, FieldDescriptions, InputField, InvocationContext, invocation
+from .baseinvocation import BaseInvocation, FieldDescriptions, Input, InputField, InvocationContext, invocation
 
 
 @invocation("show_image", title="Show Image", tags=["image"], category="image", version="1.0.0")
@@ -972,13 +972,14 @@ class ImageChannelMultiplyInvocation(BaseInvocation):
     title="Save Image",
     tags=["primitives", "image"],
     category="primitives",
-    version="1.0.0",
+    version="1.0.1",
     use_cache=False,
 )
 class SaveImageInvocation(BaseInvocation):
     """Saves an image. Unlike an image primitive, this invocation stores a copy of the image."""
 
-    image: ImageField = InputField(description="The image to load")
+    image: ImageField = InputField(description=FieldDescriptions.image)
+    board: Optional[BoardField] = InputField(default=None, description=FieldDescriptions.board, input=Input.Direct)
     metadata: CoreMetadata = InputField(
         default=None,
         description=FieldDescriptions.core_metadata,
@@ -992,6 +993,7 @@ class SaveImageInvocation(BaseInvocation):
             image=image,
             image_origin=ResourceOrigin.INTERNAL,
             image_category=ImageCategory.GENERAL,
+            board_id=self.board.board_id if self.board else None,
             node_id=self.id,
             session_id=context.graph_execution_state_id,
             is_intermediate=self.is_intermediate,

invokeai/app/invocations/infill.py

@@ -269,7 +269,7 @@ class LaMaInfillInvocation(BaseInvocation):
         )
 
 
-@invocation("infill_cv2", title="CV2 Infill", tags=["image", "inpaint"], category="inpaint")
+@invocation("infill_cv2", title="CV2 Infill", tags=["image", "inpaint"], category="inpaint", version="1.0.0")
 class CV2InfillInvocation(BaseInvocation):
     """Infills transparent areas of an image using OpenCV Inpainting"""
 

invokeai/app/invocations/ip_adapter.py

@@ -0,0 +1,103 @@
+import os
+from builtins import float
+from typing import List, Union
+
+from pydantic import BaseModel, Field
+
+from invokeai.app.invocations.baseinvocation import (
+    BaseInvocation,
+    BaseInvocationOutput,
+    FieldDescriptions,
+    Input,
+    InputField,
+    InvocationContext,
+    OutputField,
+    UIType,
+    invocation,
+    invocation_output,
+)
+from invokeai.app.invocations.primitives import ImageField
+from invokeai.backend.model_management.models.base import BaseModelType, ModelType
+from invokeai.backend.model_management.models.ip_adapter import get_ip_adapter_image_encoder_model_id
+
+
+class IPAdapterModelField(BaseModel):
+    model_name: str = Field(description="Name of the IP-Adapter model")
+    base_model: BaseModelType = Field(description="Base model")
+
+
+class CLIPVisionModelField(BaseModel):
+    model_name: str = Field(description="Name of the CLIP Vision image encoder model")
+    base_model: BaseModelType = Field(description="Base model (usually 'Any')")
+
+
+class IPAdapterField(BaseModel):
+    image: ImageField = Field(description="The IP-Adapter image prompt.")
+    ip_adapter_model: IPAdapterModelField = Field(description="The IP-Adapter model to use.")
+    image_encoder_model: CLIPVisionModelField = Field(description="The name of the CLIP image encoder model.")
+    weight: Union[float, List[float]] = Field(default=1, description="The weight given to the ControlNet")
+    # weight: float = Field(default=1.0, ge=0, description="The weight of the IP-Adapter.")
+    begin_step_percent: float = Field(
+        default=0, ge=0, le=1, description="When the IP-Adapter is first applied (% of total steps)"
+    )
+    end_step_percent: float = Field(
+        default=1, ge=0, le=1, description="When the IP-Adapter is last applied (% of total steps)"
+    )
+
+
+@invocation_output("ip_adapter_output")
+class IPAdapterOutput(BaseInvocationOutput):
+    # Outputs
+    ip_adapter: IPAdapterField = OutputField(description=FieldDescriptions.ip_adapter, title="IP-Adapter")
+
+
+@invocation("ip_adapter", title="IP-Adapter", tags=["ip_adapter", "control"], category="ip_adapter", version="1.0.0")
+class IPAdapterInvocation(BaseInvocation):
+    """Collects IP-Adapter info to pass to other nodes."""
+
+    # Inputs
+    image: ImageField = InputField(description="The IP-Adapter image prompt.")
+    ip_adapter_model: IPAdapterModelField = InputField(
+        description="The IP-Adapter model.", title="IP-Adapter Model", input=Input.Direct, ui_order=-1
+    )
+
+    # weight: float = InputField(default=1.0, description="The weight of the IP-Adapter.", ui_type=UIType.Float)
+    weight: Union[float, List[float]] = InputField(
+        default=1, ge=0, description="The weight given to the IP-Adapter", ui_type=UIType.Float, title="Weight"
+    )
+
+    begin_step_percent: float = InputField(
+        default=0, ge=-1, le=2, description="When the IP-Adapter is first applied (% of total steps)"
+    )
+    end_step_percent: float = InputField(
+        default=1, ge=0, le=1, description="When the IP-Adapter is last applied (% of total steps)"
+    )
+
+    def invoke(self, context: InvocationContext) -> IPAdapterOutput:
+        # Lookup the CLIP Vision encoder that is intended to be used with the IP-Adapter model.
+        ip_adapter_info = context.services.model_manager.model_info(
+            self.ip_adapter_model.model_name, self.ip_adapter_model.base_model, ModelType.IPAdapter
+        )
+        # HACK(ryand): This is bad for a couple of reasons: 1) we are bypassing the model manager to read the model
+        # directly, and 2) we are reading from disk every time this invocation is called without caching the result.
+        # A better solution would be to store the image encoder model reference in the IP-Adapter model info, but this
+        # is currently messy due to differences between how the model info is generated when installing a model from
+        # disk vs. downloading the model.
+        image_encoder_model_id = get_ip_adapter_image_encoder_model_id(
+            os.path.join(context.services.configuration.get_config().models_path, ip_adapter_info["path"])
+        )
+        image_encoder_model_name = image_encoder_model_id.split("/")[-1].strip()
+        image_encoder_model = CLIPVisionModelField(
+            model_name=image_encoder_model_name,
+            base_model=BaseModelType.Any,
+        )
+        return IPAdapterOutput(
+            ip_adapter=IPAdapterField(
+                image=self.image,
+                ip_adapter_model=self.ip_adapter_model,
+                image_encoder_model=image_encoder_model,
+                weight=self.weight,
+                begin_step_percent=self.begin_step_percent,
+                end_step_percent=self.end_step_percent,
+            ),
+        )

invokeai/app/invocations/latent.py

@@ -1,13 +1,16 @@
 # Copyright (c) 2023 Kyle Schouviller (https://github.com/kyle0654)
 
 from contextlib import ExitStack
+from functools import singledispatchmethod
 from typing import List, Literal, Optional, Union
 
 import einops
 import numpy as np
 import torch
 import torchvision.transforms as T
+from diffusers import AutoencoderKL, AutoencoderTiny
 from diffusers.image_processor import VaeImageProcessor
+from diffusers.models.adapter import FullAdapterXL, T2IAdapter
 from diffusers.models.attention_processor import (
     AttnProcessor2_0,
     LoRAAttnProcessor2_0,
@@ -19,6 +22,7 @@ from diffusers.schedulers import SchedulerMixin as Scheduler
 from pydantic import validator
 from torchvision.transforms.functional import resize as tv_resize
 
+from invokeai.app.invocations.ip_adapter import IPAdapterField
 from invokeai.app.invocations.metadata import CoreMetadata
 from invokeai.app.invocations.primitives import (
     DenoiseMaskField,
@@ -29,18 +33,22 @@ from invokeai.app.invocations.primitives import (
     LatentsOutput,
     build_latents_output,
 )
+from invokeai.app.invocations.t2i_adapter import T2IAdapterField
 from invokeai.app.util.controlnet_utils import prepare_control_image
 from invokeai.app.util.step_callback import stable_diffusion_step_callback
+from invokeai.backend.ip_adapter.ip_adapter import IPAdapter, IPAdapterPlus
 from invokeai.backend.model_management.models import ModelType, SilenceWarnings
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningData, IPAdapterConditioningInfo
 
 from ...backend.model_management.lora import ModelPatcher
 from ...backend.model_management.models import BaseModelType
 from ...backend.model_management.seamless import set_seamless
 from ...backend.stable_diffusion import PipelineIntermediateState
 from ...backend.stable_diffusion.diffusers_pipeline import (
-    ConditioningData,
     ControlNetData,
+    IPAdapterData,
     StableDiffusionGeneratorPipeline,
+    T2IAdapterData,
     image_resized_to_grid_as_tensor,
 )
 from ...backend.stable_diffusion.diffusion.shared_invokeai_diffusion import PostprocessingSettings
@@ -68,7 +76,6 @@ if choose_torch_device() == torch.device("mps"):
 
 DEFAULT_PRECISION = choose_precision(choose_torch_device())
 
-
 SAMPLER_NAME_VALUES = Literal[tuple(list(SCHEDULER_MAP.keys()))]
 
 
@@ -191,7 +198,7 @@ def get_scheduler(
     title="Denoise Latents",
     tags=["latents", "denoise", "txt2img", "t2i", "t2l", "img2img", "i2i", "l2l"],
     category="latents",
-    version="1.0.0",
+    version="1.3.0",
 )
 class DenoiseLatentsInvocation(BaseInvocation):
     """Denoises noisy latents to decodable images"""
@@ -205,7 +212,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
     noise: Optional[LatentsField] = InputField(description=FieldDescriptions.noise, input=Input.Connection, ui_order=3)
     steps: int = InputField(default=10, gt=0, description=FieldDescriptions.steps)
     cfg_scale: Union[float, List[float]] = InputField(
-        default=7.5, ge=1, description=FieldDescriptions.cfg_scale, ui_type=UIType.Float, title="CFG Scale"
+        default=7.5, ge=1, description=FieldDescriptions.cfg_scale, title="CFG Scale"
     )
     denoising_start: float = InputField(default=0.0, ge=0, le=1, description=FieldDescriptions.denoising_start)
     denoising_end: float = InputField(default=1.0, ge=0, le=1, description=FieldDescriptions.denoising_end)
@@ -215,13 +222,18 @@ class DenoiseLatentsInvocation(BaseInvocation):
     unet: UNetField = InputField(description=FieldDescriptions.unet, input=Input.Connection, title="UNet", ui_order=2)
     control: Union[ControlField, list[ControlField]] = InputField(
         default=None,
-        description=FieldDescriptions.control,
         input=Input.Connection,
         ui_order=5,
     )
+    ip_adapter: Optional[Union[IPAdapterField, list[IPAdapterField]]] = InputField(
+        description=FieldDescriptions.ip_adapter, title="IP-Adapter", default=None, input=Input.Connection, ui_order=6
+    )
+    t2i_adapter: Union[T2IAdapterField, list[T2IAdapterField]] = InputField(
+        description=FieldDescriptions.t2i_adapter, title="T2I-Adapter", default=None, input=Input.Connection, ui_order=7
+    )
     latents: Optional[LatentsField] = InputField(description=FieldDescriptions.latents, input=Input.Connection)
     denoise_mask: Optional[DenoiseMaskField] = InputField(
-        default=None, description=FieldDescriptions.mask, input=Input.Connection, ui_order=6
+        default=None, description=FieldDescriptions.mask, input=Input.Connection, ui_order=8
     )
 
     @validator("cfg_scale")
@@ -323,8 +335,6 @@ class DenoiseLatentsInvocation(BaseInvocation):
     def prep_control_data(
         self,
         context: InvocationContext,
-        # really only need model for dtype and device
-        model: StableDiffusionGeneratorPipeline,
         control_input: Union[ControlField, List[ControlField]],
         latents_shape: List[int],
         exit_stack: ExitStack,
@@ -344,57 +354,205 @@ class DenoiseLatentsInvocation(BaseInvocation):
         else:
             control_list = None
         if control_list is None:
-            control_data = None
-            # from above handling, any control that is not None should now be of type list[ControlField]
-        else:
-            # FIXME: add checks to skip entry if model or image is None
-            #        and if weight is None, populate with default 1.0?
-            control_data = []
-            control_models = []
-            for control_info in control_list:
-                control_model = exit_stack.enter_context(
-                    context.services.model_manager.get_model(
-                        model_name=control_info.control_model.model_name,
-                        model_type=ModelType.ControlNet,
-                        base_model=control_info.control_model.base_model,
-                        context=context,
-                    )
+            return None
+        # After above handling, any control that is not None should now be of type list[ControlField].
+
+        # FIXME: add checks to skip entry if model or image is None
+        #        and if weight is None, populate with default 1.0?
+        controlnet_data = []
+        for control_info in control_list:
+            control_model = exit_stack.enter_context(
+                context.services.model_manager.get_model(
+                    model_name=control_info.control_model.model_name,
+                    model_type=ModelType.ControlNet,
+                    base_model=control_info.control_model.base_model,
+                    context=context,
+                )
+            )
+
+            # control_models.append(control_model)
+            control_image_field = control_info.image
+            input_image = context.services.images.get_pil_image(control_image_field.image_name)
+            # self.image.image_type, self.image.image_name
+            # FIXME: still need to test with different widths, heights, devices, dtypes
+            #        and add in batch_size, num_images_per_prompt?
+            #        and do real check for classifier_free_guidance?
+            # prepare_control_image should return torch.Tensor of shape(batch_size, 3, height, width)
+            control_image = prepare_control_image(
+                image=input_image,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                width=control_width_resize,
+                height=control_height_resize,
+                # batch_size=batch_size * num_images_per_prompt,
+                # num_images_per_prompt=num_images_per_prompt,
+                device=control_model.device,
+                dtype=control_model.dtype,
+                control_mode=control_info.control_mode,
+                resize_mode=control_info.resize_mode,
+            )
+            control_item = ControlNetData(
+                model=control_model,  # model object
+                image_tensor=control_image,
+                weight=control_info.control_weight,
+                begin_step_percent=control_info.begin_step_percent,
+                end_step_percent=control_info.end_step_percent,
+                control_mode=control_info.control_mode,
+                # any resizing needed should currently be happening in prepare_control_image(),
+                #    but adding resize_mode to ControlNetData in case needed in the future
+                resize_mode=control_info.resize_mode,
+            )
+            controlnet_data.append(control_item)
+            # MultiControlNetModel has been refactored out, just need list[ControlNetData]
+
+        return controlnet_data
+
+    def prep_ip_adapter_data(
+        self,
+        context: InvocationContext,
+        ip_adapter: Optional[Union[IPAdapterField, list[IPAdapterField]]],
+        conditioning_data: ConditioningData,
+        exit_stack: ExitStack,
+    ) -> Optional[list[IPAdapterData]]:
+        """If IP-Adapter is enabled, then this function loads the requisite models, and adds the image prompt embeddings
+        to the `conditioning_data` (in-place).
+        """
+        if ip_adapter is None:
+            return None
+
+        # ip_adapter could be a list or a single IPAdapterField. Normalize to a list here.
+        if not isinstance(ip_adapter, list):
+            ip_adapter = [ip_adapter]
+
+        if len(ip_adapter) == 0:
+            return None
+
+        ip_adapter_data_list = []
+        conditioning_data.ip_adapter_conditioning = []
+        for single_ip_adapter in ip_adapter:
+            ip_adapter_model: Union[IPAdapter, IPAdapterPlus] = exit_stack.enter_context(
+                context.services.model_manager.get_model(
+                    model_name=single_ip_adapter.ip_adapter_model.model_name,
+                    model_type=ModelType.IPAdapter,
+                    base_model=single_ip_adapter.ip_adapter_model.base_model,
+                    context=context,
+                )
+            )
+
+            image_encoder_model_info = context.services.model_manager.get_model(
+                model_name=single_ip_adapter.image_encoder_model.model_name,
+                model_type=ModelType.CLIPVision,
+                base_model=single_ip_adapter.image_encoder_model.base_model,
+                context=context,
+            )
+
+            input_image = context.services.images.get_pil_image(single_ip_adapter.image.image_name)
+
+            # TODO(ryand): With some effort, the step of running the CLIP Vision encoder could be done before any other
+            # models are needed in memory. This would help to reduce peak memory utilization in low-memory environments.
+            with image_encoder_model_info as image_encoder_model:
+                # Get image embeddings from CLIP and ImageProjModel.
+                image_prompt_embeds, uncond_image_prompt_embeds = ip_adapter_model.get_image_embeds(
+                    input_image, image_encoder_model
+                )
+                conditioning_data.ip_adapter_conditioning.append(
+                    IPAdapterConditioningInfo(image_prompt_embeds, uncond_image_prompt_embeds)
                 )
 
-                control_models.append(control_model)
-                control_image_field = control_info.image
-                input_image = context.services.images.get_pil_image(control_image_field.image_name)
-                # self.image.image_type, self.image.image_name
-                # FIXME: still need to test with different widths, heights, devices, dtypes
-                #        and add in batch_size, num_images_per_prompt?
-                #        and do real check for classifier_free_guidance?
-                # prepare_control_image should return torch.Tensor of shape(batch_size, 3, height, width)
-                control_image = prepare_control_image(
-                    image=input_image,
-                    do_classifier_free_guidance=do_classifier_free_guidance,
-                    width=control_width_resize,
-                    height=control_height_resize,
-                    # batch_size=batch_size * num_images_per_prompt,
-                    # num_images_per_prompt=num_images_per_prompt,
-                    device=control_model.device,
-                    dtype=control_model.dtype,
-                    control_mode=control_info.control_mode,
-                    resize_mode=control_info.resize_mode,
+            ip_adapter_data_list.append(
+                IPAdapterData(
+                    ip_adapter_model=ip_adapter_model,
+                    weight=single_ip_adapter.weight,
+                    begin_step_percent=single_ip_adapter.begin_step_percent,
+                    end_step_percent=single_ip_adapter.end_step_percent,
                 )
-                control_item = ControlNetData(
-                    model=control_model,
-                    image_tensor=control_image,
-                    weight=control_info.control_weight,
-                    begin_step_percent=control_info.begin_step_percent,
-                    end_step_percent=control_info.end_step_percent,
-                    control_mode=control_info.control_mode,
-                    # any resizing needed should currently be happening in prepare_control_image(),
-                    #    but adding resize_mode to ControlNetData in case needed in the future
-                    resize_mode=control_info.resize_mode,
+            )
+
+        return ip_adapter_data_list
+
+    def run_t2i_adapters(
+        self,
+        context: InvocationContext,
+        t2i_adapter: Optional[Union[T2IAdapterField, list[T2IAdapterField]]],
+        latents_shape: list[int],
+        do_classifier_free_guidance: bool,
+    ) -> Optional[list[T2IAdapterData]]:
+        if t2i_adapter is None:
+            return None
+
+        # Handle the possibility that t2i_adapter could be a list or a single T2IAdapterField.
+        if isinstance(t2i_adapter, T2IAdapterField):
+            t2i_adapter = [t2i_adapter]
+
+        if len(t2i_adapter) == 0:
+            return None
+
+        t2i_adapter_data = []
+        for t2i_adapter_field in t2i_adapter:
+            t2i_adapter_model_info = context.services.model_manager.get_model(
+                model_name=t2i_adapter_field.t2i_adapter_model.model_name,
+                model_type=ModelType.T2IAdapter,
+                base_model=t2i_adapter_field.t2i_adapter_model.base_model,
+                context=context,
+            )
+            image = context.services.images.get_pil_image(t2i_adapter_field.image.image_name)
+
+            # The max_unet_downscale is the maximum amount that the UNet model downscales the latent image internally.
+            if t2i_adapter_field.t2i_adapter_model.base_model == BaseModelType.StableDiffusion1:
+                max_unet_downscale = 8
+            elif t2i_adapter_field.t2i_adapter_model.base_model == BaseModelType.StableDiffusionXL:
+                max_unet_downscale = 4
+            else:
+                raise ValueError(
+                    f"Unexpected T2I-Adapter base model type: '{t2i_adapter_field.t2i_adapter_model.base_model}'."
                 )
-                control_data.append(control_item)
-                # MultiControlNetModel has been refactored out, just need list[ControlNetData]
-        return control_data
+
+            t2i_adapter_model: T2IAdapter
+            with t2i_adapter_model_info as t2i_adapter_model:
+                total_downscale_factor = t2i_adapter_model.total_downscale_factor
+                if isinstance(t2i_adapter_model.adapter, FullAdapterXL):
+                    # HACK(ryand): Work around a bug in FullAdapterXL. This is being addressed upstream in diffusers by
+                    # this PR: https://github.com/huggingface/diffusers/pull/5134.
+                    total_downscale_factor = total_downscale_factor // 2
+
+                # Resize the T2I-Adapter input image.
+                # We select the resize dimensions so that after the T2I-Adapter's total_downscale_factor is applied, the
+                # result will match the latent image's dimensions after max_unet_downscale is applied.
+                t2i_input_height = latents_shape[2] // max_unet_downscale * total_downscale_factor
+                t2i_input_width = latents_shape[3] // max_unet_downscale * total_downscale_factor
+
+                # Note: We have hard-coded `do_classifier_free_guidance=False`. This is because we only want to prepare
+                # a single image. If CFG is enabled, we will duplicate the resultant tensor after applying the
+                # T2I-Adapter model.
+                #
+                # Note: We re-use the `prepare_control_image(...)` from ControlNet for T2I-Adapter, because it has many
+                # of the same requirements (e.g. preserving binary masks during resize).
+                t2i_image = prepare_control_image(
+                    image=image,
+                    do_classifier_free_guidance=False,
+                    width=t2i_input_width,
+                    height=t2i_input_height,
+                    num_channels=t2i_adapter_model.config.in_channels,
+                    device=t2i_adapter_model.device,
+                    dtype=t2i_adapter_model.dtype,
+                    resize_mode=t2i_adapter_field.resize_mode,
+                )
+
+                adapter_state = t2i_adapter_model(t2i_image)
+
+            if do_classifier_free_guidance:
+                for idx, value in enumerate(adapter_state):
+                    adapter_state[idx] = torch.cat([value] * 2, dim=0)
+
+            t2i_adapter_data.append(
+                T2IAdapterData(
+                    adapter_state=adapter_state,
+                    weight=t2i_adapter_field.weight,
+                    begin_step_percent=t2i_adapter_field.begin_step_percent,
+                    end_step_percent=t2i_adapter_field.end_step_percent,
+                )
+            )
+
+        return t2i_adapter_data
 
     # original idea by https://github.com/AmericanPresidentJimmyCarter
     # TODO: research more for second order schedulers timesteps
@@ -467,6 +625,12 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
             mask, masked_latents = self.prep_inpaint_mask(context, latents)
 
+            # TODO(ryand): I have hard-coded `do_classifier_free_guidance=True` to mirror the behaviour of ControlNets,
+            # below. Investigate whether this is appropriate.
+            t2i_adapter_data = self.run_t2i_adapters(
+                context, self.t2i_adapter, latents.shape, do_classifier_free_guidance=True
+            )
+
             # Get the source node id (we are invoking the prepared node)
             graph_execution_state = context.services.graph_execution_manager.get(context.graph_execution_state_id)
             source_node_id = graph_execution_state.prepared_source_mapping[self.id]
@@ -488,9 +652,12 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 **self.unet.unet.dict(),
                 context=context,
             )
-            with ExitStack() as exit_stack, ModelPatcher.apply_lora_unet(
-                unet_info.context.model, _lora_loader()
-            ), set_seamless(unet_info.context.model, self.unet.seamless_axes), unet_info as unet:
+            with (
+                ExitStack() as exit_stack,
+                ModelPatcher.apply_lora_unet(unet_info.context.model, _lora_loader()),
+                set_seamless(unet_info.context.model, self.unet.seamless_axes),
+                unet_info as unet,
+            ):
                 latents = latents.to(device=unet.device, dtype=unet.dtype)
                 if noise is not None:
                     noise = noise.to(device=unet.device, dtype=unet.dtype)
@@ -509,8 +676,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 pipeline = self.create_pipeline(unet, scheduler)
                 conditioning_data = self.get_conditioning_data(context, scheduler, unet, seed)
 
-                control_data = self.prep_control_data(
-                    model=pipeline,
+                controlnet_data = self.prep_control_data(
                     context=context,
                     control_input=self.control,
                     latents_shape=latents.shape,
@@ -519,6 +685,13 @@ class DenoiseLatentsInvocation(BaseInvocation):
                     exit_stack=exit_stack,
                 )
 
+                ip_adapter_data = self.prep_ip_adapter_data(
+                    context=context,
+                    ip_adapter=self.ip_adapter,
+                    conditioning_data=conditioning_data,
+                    exit_stack=exit_stack,
+                )
+
                 num_inference_steps, timesteps, init_timestep = self.init_scheduler(
                     scheduler,
                     device=unet.device,
@@ -537,7 +710,9 @@ class DenoiseLatentsInvocation(BaseInvocation):
                     masked_latents=masked_latents,
                     num_inference_steps=num_inference_steps,
                     conditioning_data=conditioning_data,
-                    control_data=control_data,  # list[ControlNetData]
+                    control_data=controlnet_data,
+                    ip_adapter_data=ip_adapter_data,
+                    t2i_adapter_data=t2i_adapter_data,
                     callback=step_callback,
                 )
 
@@ -792,8 +967,7 @@ class ImageToLatentsInvocation(BaseInvocation):
             # non_noised_latents_from_image
             image_tensor = image_tensor.to(device=vae.device, dtype=vae.dtype)
             with torch.inference_mode():
-                image_tensor_dist = vae.encode(image_tensor).latent_dist
-                latents = image_tensor_dist.sample().to(dtype=vae.dtype)  # FIXME: uses torch.randn. make reproducible!
+                latents = ImageToLatentsInvocation._encode_to_tensor(vae, image_tensor)
 
             latents = vae.config.scaling_factor * latents
             latents = latents.to(dtype=orig_dtype)
@@ -820,6 +994,18 @@ class ImageToLatentsInvocation(BaseInvocation):
         context.services.latents.save(name, latents)
         return build_latents_output(latents_name=name, latents=latents, seed=None)
 
+    @singledispatchmethod
+    @staticmethod
+    def _encode_to_tensor(vae: AutoencoderKL, image_tensor: torch.FloatTensor) -> torch.FloatTensor:
+        image_tensor_dist = vae.encode(image_tensor).latent_dist
+        latents = image_tensor_dist.sample().to(dtype=vae.dtype)  # FIXME: uses torch.randn. make reproducible!
+        return latents
+
+    @_encode_to_tensor.register
+    @staticmethod
+    def _(vae: AutoencoderTiny, image_tensor: torch.FloatTensor) -> torch.FloatTensor:
+        return vae.encode(image_tensor).latents
+
 
 @invocation("lblend", title="Blend Latents", tags=["latents", "blend"], category="latents", version="1.0.0")
 class BlendLatentsInvocation(BaseInvocation):

invokeai/app/invocations/math.py

@@ -65,13 +65,27 @@ class DivideInvocation(BaseInvocation):
 class RandomIntInvocation(BaseInvocation):
     """Outputs a single random integer."""
 
-    low: int = InputField(default=0, description="The inclusive low value")
-    high: int = InputField(default=np.iinfo(np.int32).max, description="The exclusive high value")
+    low: int = InputField(default=0, description=FieldDescriptions.inclusive_low)
+    high: int = InputField(default=np.iinfo(np.int32).max, description=FieldDescriptions.exclusive_high)
 
     def invoke(self, context: InvocationContext) -> IntegerOutput:
         return IntegerOutput(value=np.random.randint(self.low, self.high))
 
 
+@invocation("rand_float", title="Random Float", tags=["math", "float", "random"], category="math", version="1.0.0")
+class RandomFloatInvocation(BaseInvocation):
+    """Outputs a single random float"""
+
+    low: float = InputField(default=0.0, description=FieldDescriptions.inclusive_low)
+    high: float = InputField(default=1.0, description=FieldDescriptions.exclusive_high)
+    decimals: int = InputField(default=2, description=FieldDescriptions.decimal_places)
+
+    def invoke(self, context: InvocationContext) -> FloatOutput:
+        random_float = np.random.uniform(self.low, self.high)
+        rounded_float = round(random_float, self.decimals)
+        return FloatOutput(value=rounded_float)
+
+
 @invocation(
     "float_to_int",
     title="Float To Integer",

invokeai/app/invocations/metadata.py

@@ -12,7 +12,10 @@ from invokeai.app.invocations.baseinvocation import (
     invocation_output,
 )
 from invokeai.app.invocations.controlnet_image_processors import ControlField
+from invokeai.app.invocations.ip_adapter import IPAdapterModelField
 from invokeai.app.invocations.model import LoRAModelField, MainModelField, VAEModelField
+from invokeai.app.invocations.primitives import ImageField
+from invokeai.app.invocations.t2i_adapter import T2IAdapterField
 from invokeai.app.util.model_exclude_null import BaseModelExcludeNull
 
 from ...version import __version__
@@ -25,6 +28,18 @@ class LoRAMetadataField(BaseModelExcludeNull):
     weight: float = Field(description="The weight of the LoRA model")
 
 
+class IPAdapterMetadataField(BaseModelExcludeNull):
+    image: ImageField = Field(description="The IP-Adapter image prompt.")
+    ip_adapter_model: IPAdapterModelField = Field(description="The IP-Adapter model to use.")
+    weight: float = Field(description="The weight of the IP-Adapter model")
+    begin_step_percent: float = Field(
+        default=0, ge=0, le=1, description="When the IP-Adapter is first applied (% of total steps)"
+    )
+    end_step_percent: float = Field(
+        default=1, ge=0, le=1, description="When the IP-Adapter is last applied (% of total steps)"
+    )
+
+
 class CoreMetadata(BaseModelExcludeNull):
     """Core generation metadata for an image generated in InvokeAI."""
 
@@ -42,11 +57,14 @@ class CoreMetadata(BaseModelExcludeNull):
     cfg_scale: float = Field(description="The classifier-free guidance scale parameter")
     steps: int = Field(description="The number of steps used for inference")
     scheduler: str = Field(description="The scheduler used for inference")
-    clip_skip: int = Field(
+    clip_skip: Optional[int] = Field(
+        default=None,
         description="The number of skipped CLIP layers",
     )
     model: MainModelField = Field(description="The main model used for inference")
     controlnets: list[ControlField] = Field(description="The ControlNets used for inference")
+    ipAdapters: list[IPAdapterMetadataField] = Field(description="The IP Adapters used for inference")
+    t2iAdapters: list[T2IAdapterField] = Field(description="The IP Adapters used for inference")
     loras: list[LoRAMetadataField] = Field(description="The LoRAs used for inference")
     vae: Optional[VAEModelField] = Field(
         default=None,
@@ -116,11 +134,14 @@ class MetadataAccumulatorInvocation(BaseInvocation):
     cfg_scale: float = InputField(description="The classifier-free guidance scale parameter")
     steps: int = InputField(description="The number of steps used for inference")
     scheduler: str = InputField(description="The scheduler used for inference")
-    clip_skip: int = InputField(
+    clip_skip: Optional[int] = Field(
+        default=None,
         description="The number of skipped CLIP layers",
     )
     model: MainModelField = InputField(description="The main model used for inference")
     controlnets: list[ControlField] = InputField(description="The ControlNets used for inference")
+    ipAdapters: list[IPAdapterMetadataField] = InputField(description="The IP Adapters used for inference")
+    t2iAdapters: list[T2IAdapterField] = Field(description="The IP Adapters used for inference")
     loras: list[LoRAMetadataField] = InputField(description="The LoRAs used for inference")
     strength: Optional[float] = InputField(
         default=None,

invokeai/app/invocations/onnx.py

@@ -95,9 +95,10 @@ class ONNXPromptInvocation(BaseInvocation):
                     print(f'Warn: trigger: "{trigger}" not found')
             if loras or ti_list:
                 text_encoder.release_session()
-            with ONNXModelPatcher.apply_lora_text_encoder(text_encoder, loras), ONNXModelPatcher.apply_ti(
-                orig_tokenizer, text_encoder, ti_list
-            ) as (tokenizer, ti_manager):
+            with (
+                ONNXModelPatcher.apply_lora_text_encoder(text_encoder, loras),
+                ONNXModelPatcher.apply_ti(orig_tokenizer, text_encoder, ti_list) as (tokenizer, ti_manager),
+            ):
                 text_encoder.create_session()
 
                 # copy from
@@ -165,7 +166,6 @@ class ONNXTextToLatentsInvocation(BaseInvocation):
         default=7.5,
         ge=1,
         description=FieldDescriptions.cfg_scale,
-        ui_type=UIType.Float,
     )
     scheduler: SAMPLER_NAME_VALUES = InputField(
         default="euler", description=FieldDescriptions.scheduler, input=Input.Direct, ui_type=UIType.Scheduler
@@ -178,7 +178,6 @@ class ONNXTextToLatentsInvocation(BaseInvocation):
     control: Optional[Union[ControlField, list[ControlField]]] = InputField(
         default=None,
         description=FieldDescriptions.control,
-        ui_type=UIType.Control,
     )
     # seamless:   bool = InputField(default=False, description="Whether or not to generate an image that can tile without seams", )
     # seamless_axes: str = InputField(default="", description="The axes to tile the image on, 'x' and/or 'y'")

invokeai/app/invocations/primitives.py

@@ -226,6 +226,12 @@ class ImageField(BaseModel):
     image_name: str = Field(description="The name of the image")
 
 
+class BoardField(BaseModel):
+    """A board primitive field"""
+
+    board_id: str = Field(description="The id of the board")
+
+
 @invocation_output("image_output")
 class ImageOutput(BaseInvocationOutput):
     """Base class for nodes that output a single image"""

invokeai/app/invocations/t2i_adapter.py

@@ -0,0 +1,83 @@
+from typing import Union
+
+from pydantic import BaseModel, Field
+
+from invokeai.app.invocations.baseinvocation import (
+    BaseInvocation,
+    BaseInvocationOutput,
+    FieldDescriptions,
+    Input,
+    InputField,
+    InvocationContext,
+    OutputField,
+    UIType,
+    invocation,
+    invocation_output,
+)
+from invokeai.app.invocations.controlnet_image_processors import CONTROLNET_RESIZE_VALUES
+from invokeai.app.invocations.primitives import ImageField
+from invokeai.backend.model_management.models.base import BaseModelType
+
+
+class T2IAdapterModelField(BaseModel):
+    model_name: str = Field(description="Name of the T2I-Adapter model")
+    base_model: BaseModelType = Field(description="Base model")
+
+
+class T2IAdapterField(BaseModel):
+    image: ImageField = Field(description="The T2I-Adapter image prompt.")
+    t2i_adapter_model: T2IAdapterModelField = Field(description="The T2I-Adapter model to use.")
+    weight: Union[float, list[float]] = Field(default=1, description="The weight given to the T2I-Adapter")
+    begin_step_percent: float = Field(
+        default=0, ge=0, le=1, description="When the T2I-Adapter is first applied (% of total steps)"
+    )
+    end_step_percent: float = Field(
+        default=1, ge=0, le=1, description="When the T2I-Adapter is last applied (% of total steps)"
+    )
+    resize_mode: CONTROLNET_RESIZE_VALUES = Field(default="just_resize", description="The resize mode to use")
+
+
+@invocation_output("t2i_adapter_output")
+class T2IAdapterOutput(BaseInvocationOutput):
+    t2i_adapter: T2IAdapterField = OutputField(description=FieldDescriptions.t2i_adapter, title="T2I Adapter")
+
+
+@invocation(
+    "t2i_adapter", title="T2I-Adapter", tags=["t2i_adapter", "control"], category="t2i_adapter", version="1.0.0"
+)
+class T2IAdapterInvocation(BaseInvocation):
+    """Collects T2I-Adapter info to pass to other nodes."""
+
+    # Inputs
+    image: ImageField = InputField(description="The IP-Adapter image prompt.")
+    t2i_adapter_model: T2IAdapterModelField = InputField(
+        description="The T2I-Adapter model.",
+        title="T2I-Adapter Model",
+        input=Input.Direct,
+        ui_order=-1,
+    )
+    weight: Union[float, list[float]] = InputField(
+        default=1, ge=0, description="The weight given to the T2I-Adapter", ui_type=UIType.Float, title="Weight"
+    )
+    begin_step_percent: float = InputField(
+        default=0, ge=-1, le=2, description="When the T2I-Adapter is first applied (% of total steps)"
+    )
+    end_step_percent: float = InputField(
+        default=1, ge=0, le=1, description="When the T2I-Adapter is last applied (% of total steps)"
+    )
+    resize_mode: CONTROLNET_RESIZE_VALUES = InputField(
+        default="just_resize",
+        description="The resize mode applied to the T2I-Adapter input image so that it matches the target output size.",
+    )
+
+    def invoke(self, context: InvocationContext) -> T2IAdapterOutput:
+        return T2IAdapterOutput(
+            t2i_adapter=T2IAdapterField(
+                image=self.image,
+                t2i_adapter_model=self.t2i_adapter_model,
+                weight=self.weight,
+                begin_step_percent=self.begin_step_percent,
+                end_step_percent=self.end_step_percent,
+                resize_mode=self.resize_mode,
+            )
+        )

invokeai/app/invocations/upscale.py

@@ -4,12 +4,14 @@ from typing import Literal
 
 import cv2 as cv
 import numpy as np
+import torch
 from basicsr.archs.rrdbnet_arch import RRDBNet
 from PIL import Image
 from realesrgan import RealESRGANer
 
 from invokeai.app.invocations.primitives import ImageField, ImageOutput
 from invokeai.app.models.image import ImageCategory, ResourceOrigin
+from invokeai.backend.util.devices import choose_torch_device
 
 from .baseinvocation import BaseInvocation, InputField, InvocationContext, invocation
 
@@ -22,13 +24,19 @@ ESRGAN_MODELS = Literal[
     "RealESRGAN_x2plus.pth",
 ]
 
+if choose_torch_device() == torch.device("mps"):
+    from torch import mps
 
-@invocation("esrgan", title="Upscale (RealESRGAN)", tags=["esrgan", "upscale"], category="esrgan", version="1.0.0")
+
+@invocation("esrgan", title="Upscale (RealESRGAN)", tags=["esrgan", "upscale"], category="esrgan", version="1.1.0")
 class ESRGANInvocation(BaseInvocation):
     """Upscales an image using RealESRGAN."""
 
     image: ImageField = InputField(description="The input image")
     model_name: ESRGAN_MODELS = InputField(default="RealESRGAN_x4plus.pth", description="The Real-ESRGAN model to use")
+    tile_size: int = InputField(
+        default=400, ge=0, description="Tile size for tiled ESRGAN upscaling (0=tiling disabled)"
+    )
 
     def invoke(self, context: InvocationContext) -> ImageOutput:
         image = context.services.images.get_pil_image(self.image.image_name)
@@ -86,9 +94,11 @@ class ESRGANInvocation(BaseInvocation):
             model_path=str(models_path / esrgan_model_path),
             model=rrdbnet_model,
             half=False,
+            tile=self.tile_size,
         )
 
         # prepare image - Real-ESRGAN uses cv2 internally, and cv2 uses BGR vs RGB for PIL
+        # TODO: This strips the alpha... is that okay?
         cv_image = cv.cvtColor(np.array(image.convert("RGB")), cv.COLOR_RGB2BGR)
 
         # We can pass an `outscale` value here, but it just resizes the image by that factor after
@@ -99,6 +109,10 @@ class ESRGANInvocation(BaseInvocation):
         # back to PIL
         pil_image = Image.fromarray(cv.cvtColor(upscaled_image, cv.COLOR_BGR2RGB)).convert("RGBA")
 
+        torch.cuda.empty_cache()
+        if choose_torch_device() == torch.device("mps"):
+            mps.empty_cache()
+
         image_dto = context.services.images.create(
             image=pil_image,
             image_origin=ResourceOrigin.INTERNAL,

invokeai/app/services/config/invokeai_config.py

@@ -241,8 +241,8 @@ class InvokeAIAppConfig(InvokeAISettings):
     version             : bool = Field(default=False, description="Show InvokeAI version and exit", category="Other")
 
     # CACHE
-    ram                 : Union[float, Literal["auto"]] = Field(default=6.0, gt=0, description="Maximum memory amount used by model cache for rapid switching (floating point number or 'auto')", category="Model Cache", )
-    vram                : Union[float, Literal["auto"]] = Field(default=0.25, ge=0, description="Amount of VRAM reserved for model storage (floating point number or 'auto')", category="Model Cache", )
+    ram                 : float = Field(default=7.5, gt=0, description="Maximum memory amount used by model cache for rapid switching (floating point number, GB)", category="Model Cache", )
+    vram                : float = Field(default=0.25, ge=0, description="Amount of VRAM reserved for model storage (floating point number, GB)", category="Model Cache", )
     lazy_offload        : bool = Field(default=True, description="Keep models in VRAM until their space is needed", category="Model Cache", )
 
     # DEVICE
@@ -255,6 +255,7 @@ class InvokeAIAppConfig(InvokeAISettings):
     attention_slice_size: Literal["auto", "balanced", "max", 1, 2, 3, 4, 5, 6, 7, 8] = Field(default="auto", description='Slice size, valid when attention_type=="sliced"', category="Generation", )
     force_tiled_decode  : bool = Field(default=False, description="Whether to enable tiled VAE decode (reduces memory consumption with some performance penalty)", category="Generation",)
     force_tiled_decode: bool = Field(default=False, description="Whether to enable tiled VAE decode (reduces memory consumption with some performance penalty)", category="Generation",)
+    png_compress_level  : int = Field(default=6, description="The compress_level setting of PIL.Image.save(), used for PNG encoding. All settings are lossless. 0 = fastest, largest filesize, 9 = slowest, smallest filesize", category="Generation", )
 
     # QUEUE
     max_queue_size      : int = Field(default=10000, gt=0, description="Maximum number of items in the session queue", category="Queue", )
@@ -277,6 +278,7 @@ class InvokeAIAppConfig(InvokeAISettings):
 
     class Config:
         validate_assignment = True
+        env_prefix = "INVOKEAI"
 
     def parse_args(self, argv: Optional[list[str]] = None, conf: Optional[DictConfig] = None, clobber=False):
         """

invokeai/app/services/events.py

@@ -4,7 +4,12 @@ from typing import Any, Optional
 
 from invokeai.app.models.image import ProgressImage
 from invokeai.app.services.model_manager_service import BaseModelType, ModelInfo, ModelType, SubModelType
-from invokeai.app.services.session_queue.session_queue_common import EnqueueBatchResult, SessionQueueItem
+from invokeai.app.services.session_queue.session_queue_common import (
+    BatchStatus,
+    EnqueueBatchResult,
+    SessionQueueItem,
+    SessionQueueStatus,
+)
 from invokeai.app.util.misc import get_timestamp
 
 
@@ -29,7 +34,8 @@ class EventServiceBase:
     def emit_generator_progress(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         node: dict,
         source_node_id: str,
@@ -44,6 +50,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 node_id=node.get("id"),
                 source_node_id=source_node_id,
@@ -57,7 +64,8 @@ class EventServiceBase:
     def emit_invocation_complete(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         result: dict,
         node: dict,
@@ -69,6 +77,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 node=node,
                 source_node_id=source_node_id,
@@ -79,7 +88,8 @@ class EventServiceBase:
     def emit_invocation_error(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         node: dict,
         source_node_id: str,
@@ -92,6 +102,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 node=node,
                 source_node_id=source_node_id,
@@ -101,7 +112,13 @@ class EventServiceBase:
         )
 
     def emit_invocation_started(
-        self, queue_id: str, queue_item_id: str, graph_execution_state_id: str, node: dict, source_node_id: str
+        self,
+        queue_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
+        graph_execution_state_id: str,
+        node: dict,
+        source_node_id: str,
     ) -> None:
         """Emitted when an invocation has started"""
         self.__emit_queue_event(
@@ -109,19 +126,23 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 node=node,
                 source_node_id=source_node_id,
             ),
         )
 
-    def emit_graph_execution_complete(self, queue_id: str, queue_item_id: str, graph_execution_state_id: str) -> None:
+    def emit_graph_execution_complete(
+        self, queue_id: str, queue_item_id: int, queue_batch_id: str, graph_execution_state_id: str
+    ) -> None:
         """Emitted when a session has completed all invocations"""
         self.__emit_queue_event(
             event_name="graph_execution_state_complete",
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
             ),
         )
@@ -129,7 +150,8 @@ class EventServiceBase:
     def emit_model_load_started(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         model_name: str,
         base_model: BaseModelType,
@@ -142,6 +164,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 model_name=model_name,
                 base_model=base_model,
@@ -153,7 +176,8 @@ class EventServiceBase:
     def emit_model_load_completed(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         model_name: str,
         base_model: BaseModelType,
@@ -167,6 +191,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 model_name=model_name,
                 base_model=base_model,
@@ -181,7 +206,8 @@ class EventServiceBase:
     def emit_session_retrieval_error(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         error_type: str,
         error: str,
@@ -192,6 +218,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 error_type=error_type,
                 error=error,
@@ -201,7 +228,8 @@ class EventServiceBase:
     def emit_invocation_retrieval_error(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
         node_id: str,
         error_type: str,
@@ -213,6 +241,7 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
                 node_id=node_id,
                 error_type=error_type,
@@ -223,7 +252,8 @@ class EventServiceBase:
     def emit_session_canceled(
         self,
         queue_id: str,
-        queue_item_id: str,
+        queue_item_id: int,
+        queue_batch_id: str,
         graph_execution_state_id: str,
     ) -> None:
         """Emitted when a session is canceled"""
@@ -232,25 +262,36 @@ class EventServiceBase:
             payload=dict(
                 queue_id=queue_id,
                 queue_item_id=queue_item_id,
+                queue_batch_id=queue_batch_id,
                 graph_execution_state_id=graph_execution_state_id,
             ),
         )
 
-    def emit_queue_item_status_changed(self, session_queue_item: SessionQueueItem) -> None:
+    def emit_queue_item_status_changed(
+        self,
+        session_queue_item: SessionQueueItem,
+        batch_status: BatchStatus,
+        queue_status: SessionQueueStatus,
+    ) -> None:
         """Emitted when a queue item's status changes"""
         self.__emit_queue_event(
             event_name="queue_item_status_changed",
             payload=dict(
-                queue_id=session_queue_item.queue_id,
-                queue_item_id=session_queue_item.item_id,
-                status=session_queue_item.status,
-                batch_id=session_queue_item.batch_id,
-                session_id=session_queue_item.session_id,
-                error=session_queue_item.error,
-                created_at=str(session_queue_item.created_at) if session_queue_item.created_at else None,
-                updated_at=str(session_queue_item.updated_at) if session_queue_item.updated_at else None,
-                started_at=str(session_queue_item.started_at) if session_queue_item.started_at else None,
-                completed_at=str(session_queue_item.completed_at) if session_queue_item.completed_at else None,
+                queue_id=queue_status.queue_id,
+                queue_item=dict(
+                    queue_id=session_queue_item.queue_id,
+                    item_id=session_queue_item.item_id,
+                    status=session_queue_item.status,
+                    batch_id=session_queue_item.batch_id,
+                    session_id=session_queue_item.session_id,
+                    error=session_queue_item.error,
+                    created_at=str(session_queue_item.created_at) if session_queue_item.created_at else None,
+                    updated_at=str(session_queue_item.updated_at) if session_queue_item.updated_at else None,
+                    started_at=str(session_queue_item.started_at) if session_queue_item.started_at else None,
+                    completed_at=str(session_queue_item.completed_at) if session_queue_item.completed_at else None,
+                ),
+                batch_status=batch_status.dict(),
+                queue_status=queue_status.dict(),
             ),
         )
 

invokeai/app/services/graph.py

@@ -2,7 +2,7 @@
 
 import copy
 import itertools
-from typing import Annotated, Any, Optional, Union, cast, get_args, get_origin, get_type_hints
+from typing import Annotated, Any, Optional, Union, get_args, get_origin, get_type_hints
 
 import networkx as nx
 from pydantic import BaseModel, root_validator, validator
@@ -117,6 +117,10 @@ def are_connection_types_compatible(from_type: Any, to_type: Any) -> bool:
         if from_type is int and to_type is float:
             return True
 
+        # allow int|float -> str, pydantic will cast for us
+        if (from_type is int or from_type is float) and to_type is str:
+            return True
+
         # if not issubclass(from_type, to_type):
         if not is_union_subtype(from_type, to_type):
             return False
@@ -166,6 +170,18 @@ class NodeIdMismatchError(ValueError):
     pass
 
 
+class InvalidSubGraphError(ValueError):
+    pass
+
+
+class CyclicalGraphError(ValueError):
+    pass
+
+
+class UnknownGraphValidationError(ValueError):
+    pass
+
+
 # TODO: Create and use an Empty output?
 @invocation_output("graph_output")
 class GraphInvocationOutput(BaseInvocationOutput):
@@ -250,59 +266,6 @@ class Graph(BaseModel):
         default_factory=list,
     )
 
-    @root_validator
-    def validate_nodes_and_edges(cls, values):
-        """Validates that all edges match nodes in the graph"""
-        nodes = cast(Optional[dict[str, BaseInvocation]], values.get("nodes"))
-        edges = cast(Optional[list[Edge]], values.get("edges"))
-
-        if nodes is not None:
-            # Validate that all node ids are unique
-            node_ids = [n.id for n in nodes.values()]
-            duplicate_node_ids = set([node_id for node_id in node_ids if node_ids.count(node_id) >= 2])
-            if duplicate_node_ids:
-                raise DuplicateNodeIdError(f"Node ids must be unique, found duplicates {duplicate_node_ids}")
-
-            # Validate that all node ids match the keys in the nodes dict
-            for k, v in nodes.items():
-                if k != v.id:
-                    raise NodeIdMismatchError(f"Node ids must match, got {k} and {v.id}")
-
-        if edges is not None and nodes is not None:
-            # Validate that all edges match nodes in the graph
-            node_ids = set([e.source.node_id for e in edges] + [e.destination.node_id for e in edges])
-            missing_node_ids = [node_id for node_id in node_ids if node_id not in nodes]
-            if missing_node_ids:
-                raise NodeNotFoundError(
-                    f"All edges must reference nodes in the graph, missing nodes: {missing_node_ids}"
-                )
-
-            # Validate that all edge fields match node fields in the graph
-            for edge in edges:
-                source_node = nodes.get(edge.source.node_id, None)
-                if source_node is None:
-                    raise NodeFieldNotFoundError(f"Edge source node {edge.source.node_id} does not exist in the graph")
-
-                destination_node = nodes.get(edge.destination.node_id, None)
-                if destination_node is None:
-                    raise NodeFieldNotFoundError(
-                        f"Edge destination node {edge.destination.node_id} does not exist in the graph"
-                    )
-
-                # output fields are not on the node object directly, they are on the output type
-                if edge.source.field not in source_node.get_output_type().__fields__:
-                    raise NodeFieldNotFoundError(
-                        f"Edge source field {edge.source.field} does not exist in node {edge.source.node_id}"
-                    )
-
-                # input fields are on the node
-                if edge.destination.field not in destination_node.__fields__:
-                    raise NodeFieldNotFoundError(
-                        f"Edge destination field {edge.destination.field} does not exist in node {edge.destination.node_id}"
-                    )
-
-        return values
-
     def add_node(self, node: BaseInvocation) -> None:
         """Adds a node to a graph
 
@@ -373,53 +336,108 @@ class Graph(BaseModel):
         except KeyError:
             pass
 
-    def is_valid(self) -> bool:
-        """Validates the graph."""
+    def validate_self(self) -> None:
+        """
+        Validates the graph.
+
+        Raises an exception if the graph is invalid:
+        - `DuplicateNodeIdError`
+        - `NodeIdMismatchError`
+        - `InvalidSubGraphError`
+        - `NodeNotFoundError`
+        - `NodeFieldNotFoundError`
+        - `CyclicalGraphError`
+        - `InvalidEdgeError`
+        """
+
+        # Validate that all node ids are unique
+        node_ids = [n.id for n in self.nodes.values()]
+        duplicate_node_ids = set([node_id for node_id in node_ids if node_ids.count(node_id) >= 2])
+        if duplicate_node_ids:
+            raise DuplicateNodeIdError(f"Node ids must be unique, found duplicates {duplicate_node_ids}")
+
+        # Validate that all node ids match the keys in the nodes dict
+        for k, v in self.nodes.items():
+            if k != v.id:
+                raise NodeIdMismatchError(f"Node ids must match, got {k} and {v.id}")
 
         # Validate all subgraphs
         for gn in (n for n in self.nodes.values() if isinstance(n, GraphInvocation)):
-            if not gn.graph.is_valid():
-                return False
+            try:
+                gn.graph.validate_self()
+            except Exception as e:
+                raise InvalidSubGraphError(f"Subgraph {gn.id} is invalid") from e
 
-        # Validate all edges reference nodes in the graph
-        node_ids = set([e.source.node_id for e in self.edges] + [e.destination.node_id for e in self.edges])
-        if not all((self.has_node(node_id) for node_id in node_ids)):
-            return False
+        # Validate that all edges match nodes and fields in the graph
+        for edge in self.edges:
+            source_node = self.nodes.get(edge.source.node_id, None)
+            if source_node is None:
+                raise NodeNotFoundError(f"Edge source node {edge.source.node_id} does not exist in the graph")
+
+            destination_node = self.nodes.get(edge.destination.node_id, None)
+            if destination_node is None:
+                raise NodeNotFoundError(f"Edge destination node {edge.destination.node_id} does not exist in the graph")
+
+            # output fields are not on the node object directly, they are on the output type
+            if edge.source.field not in source_node.get_output_type().__fields__:
+                raise NodeFieldNotFoundError(
+                    f"Edge source field {edge.source.field} does not exist in node {edge.source.node_id}"
+                )
+
+            # input fields are on the node
+            if edge.destination.field not in destination_node.__fields__:
+                raise NodeFieldNotFoundError(
+                    f"Edge destination field {edge.destination.field} does not exist in node {edge.destination.node_id}"
+                )
 
         # Validate there are no cycles
         g = self.nx_graph_flat()
         if not nx.is_directed_acyclic_graph(g):
-            return False
+            raise CyclicalGraphError("Graph contains cycles")
 
         # Validate all edge connections are valid
-        if not all(
-            (
-                are_connections_compatible(
-                    self.get_node(e.source.node_id),
-                    e.source.field,
-                    self.get_node(e.destination.node_id),
-                    e.destination.field,
+        for e in self.edges:
+            if not are_connections_compatible(
+                self.get_node(e.source.node_id),
+                e.source.field,
+                self.get_node(e.destination.node_id),
+                e.destination.field,
+            ):
+                raise InvalidEdgeError(
+                    f"Invalid edge from {e.source.node_id}.{e.source.field} to {e.destination.node_id}.{e.destination.field}"
                 )
-                for e in self.edges
-            )
+
+        # Validate all iterators & collectors
+        # TODO: may need to validate all iterators & collectors in subgraphs so edge connections in parent graphs will be available
+        for n in self.nodes.values():
+            if isinstance(n, IterateInvocation) and not self._is_iterator_connection_valid(n.id):
+                raise InvalidEdgeError(f"Invalid iterator node {n.id}")
+            if isinstance(n, CollectInvocation) and not self._is_collector_connection_valid(n.id):
+                raise InvalidEdgeError(f"Invalid collector node {n.id}")
+
+        return None
+
+    def is_valid(self) -> bool:
+        """
+        Checks if the graph is valid.
+
+        Raises `UnknownGraphValidationError` if there is a problem validating the graph (not a validation error).
+        """
+        try:
+            self.validate_self()
+            return True
+        except (
+            DuplicateNodeIdError,
+            NodeIdMismatchError,
+            InvalidSubGraphError,
+            NodeNotFoundError,
+            NodeFieldNotFoundError,
+            CyclicalGraphError,
+            InvalidEdgeError,
         ):
             return False
-
-        # Validate all iterators
-        # TODO: may need to validate all iterators in subgraphs so edge connections in parent graphs will be available
-        if not all(
-            (self._is_iterator_connection_valid(n.id) for n in self.nodes.values() if isinstance(n, IterateInvocation))
-        ):
-            return False
-
-        # Validate all collectors
-        # TODO: may need to validate all collectors in subgraphs so edge connections in parent graphs will be available
-        if not all(
-            (self._is_collector_connection_valid(n.id) for n in self.nodes.values() if isinstance(n, CollectInvocation))
-        ):
-            return False
-
-        return True
+        except Exception as e:
+            raise UnknownGraphValidationError(f"Problem validating graph {e}") from e
 
     def _validate_edge(self, edge: Edge):
         """Validates that a new edge doesn't create a cycle in the graph"""
@@ -800,6 +818,12 @@ class GraphExecutionState(BaseModel):
         default_factory=dict,
     )
 
+    @validator("graph")
+    def graph_is_valid(cls, v: Graph):
+        """Validates that the graph is valid"""
+        v.validate_self()
+        return v
+
     class Config:
         schema_extra = {
             "required": [

invokeai/app/services/image_file_storage.py

@@ -9,6 +9,7 @@ from PIL import Image, PngImagePlugin
 from PIL.Image import Image as PILImageType
 from send2trash import send2trash
 
+from invokeai.app.services.config.invokeai_config import InvokeAIAppConfig
 from invokeai.app.util.thumbnails import get_thumbnail_name, make_thumbnail
 
 
@@ -79,6 +80,7 @@ class DiskImageFileStorage(ImageFileStorageBase):
     __cache_ids: Queue  # TODO: this is an incredibly naive cache
     __cache: Dict[Path, PILImageType]
     __max_cache_size: int
+    __compress_level: int
 
     def __init__(self, output_folder: Union[str, Path]):
         self.__cache = dict()
@@ -87,7 +89,7 @@ class DiskImageFileStorage(ImageFileStorageBase):
 
         self.__output_folder: Path = output_folder if isinstance(output_folder, Path) else Path(output_folder)
         self.__thumbnails_folder = self.__output_folder / "thumbnails"
-
+        self.__compress_level = InvokeAIAppConfig.get_config().png_compress_level
         # Validate required output folders at launch
         self.__validate_storage_folders()
 
@@ -134,7 +136,7 @@ class DiskImageFileStorage(ImageFileStorageBase):
                 if original_workflow is not None:
                     pnginfo.add_text("invokeai_workflow", original_workflow)
 
-            image.save(image_path, "PNG", pnginfo=pnginfo)
+            image.save(image_path, "PNG", pnginfo=pnginfo, compress_level=self.__compress_level)
 
             thumbnail_name = get_thumbnail_name(image_name)
             thumbnail_path = self.get_path(thumbnail_name, thumbnail=True)

invokeai/app/services/image_record_storage.py

@@ -584,7 +584,7 @@ class SqliteImageRecordStorage(ImageRecordStorageBase):
                 FROM images
                 JOIN board_images ON images.image_name = board_images.image_name
                 WHERE board_images.board_id = ?
-                ORDER BY images.created_at DESC
+                ORDER BY images.starred DESC, images.created_at DESC
                 LIMIT 1;
                 """,
                 (board_id,),

invokeai/app/services/images.py

@@ -41,23 +41,25 @@ class ImageServiceABC(ABC):
     _on_changed_callbacks: list[Callable[[ImageDTO], None]]
     _on_deleted_callbacks: list[Callable[[str], None]]
 
-    @abstractmethod
+    def __init__(self) -> None:
+        self._on_changed_callbacks = list()
+        self._on_deleted_callbacks = list()
+
     def on_changed(self, on_changed: Callable[[ImageDTO], None]) -> None:
-        """Register a callback for when an item is changed"""
-        pass
+        """Register a callback for when an image is changed"""
+        self._on_changed_callbacks.append(on_changed)
 
-    @abstractmethod
     def on_deleted(self, on_deleted: Callable[[str], None]) -> None:
-        """Register a callback for when an item is deleted"""
-        pass
+        """Register a callback for when an image is deleted"""
+        self._on_deleted_callbacks.append(on_deleted)
 
-    @abstractmethod
     def _on_changed(self, item: ImageDTO) -> None:
-        pass
+        for callback in self._on_changed_callbacks:
+            callback(item)
 
-    @abstractmethod
     def _on_deleted(self, item_id: str) -> None:
-        pass
+        for callback in self._on_deleted_callbacks:
+            callback(item_id)
 
     @abstractmethod
     def create(
@@ -180,26 +182,9 @@ class ImageServiceDependencies:
 
 class ImageService(ImageServiceABC):
     _services: ImageServiceDependencies
-    _on_changed_callbacks: list[Callable[[ImageDTO], None]] = list()
-    _on_deleted_callbacks: list[Callable[[str], None]] = list()
-
-    def on_changed(self, on_changed: Callable[[ImageDTO], None]) -> None:
-        """Register a callback for when an item is changed"""
-        self._on_changed_callbacks.append(on_changed)
-
-    def on_deleted(self, on_deleted: Callable[[str], None]) -> None:
-        """Register a callback for when an item is deleted"""
-        self._on_deleted_callbacks.append(on_deleted)
-
-    def _on_changed(self, item: ImageDTO) -> None:
-        for callback in self._on_changed_callbacks:
-            callback(item)
-
-    def _on_deleted(self, item_id: str) -> None:
-        for callback in self._on_deleted_callbacks:
-            callback(item_id)
 
     def __init__(self, services: ImageServiceDependencies):
+        super().__init__()
         self._services = services
 
     def create(

invokeai/app/services/invocation_cache/invocation_cache_base.py

@@ -2,24 +2,38 @@ from abc import ABC, abstractmethod
 from typing import Optional, Union
 
 from invokeai.app.invocations.baseinvocation import BaseInvocation, BaseInvocationOutput
+from invokeai.app.services.invocation_cache.invocation_cache_common import InvocationCacheStatus
 
 
 class InvocationCacheBase(ABC):
-    """Base class for invocation caches."""
+    """
+    Base class for invocation caches.
+    When an invocation is executed, it is hashed and its output stored in the cache.
+    When new invocations are executed, if they are flagged with `use_cache`, they
+    will attempt to pull their value from the cache before executing.
+
+    Implementations should register for the `on_deleted` event of the `images` and `latents`
+    services, and delete any cached outputs that reference the deleted image or latent.
+
+    See the memory implementation for an example.
+
+    Implementations should respect the `node_cache_size` configuration value, and skip all
+    cache logic if the value is set to 0.
+    """
 
     @abstractmethod
     def get(self, key: Union[int, str]) -> Optional[BaseInvocationOutput]:
-        """Retrieves and invocation output from the cache"""
+        """Retrieves an invocation output from the cache"""
         pass
 
     @abstractmethod
-    def save(self, value: BaseInvocationOutput) -> None:
+    def save(self, key: Union[int, str], invocation_output: BaseInvocationOutput) -> None:
         """Stores an invocation output in the cache"""
         pass
 
     @abstractmethod
     def delete(self, key: Union[int, str]) -> None:
-        """Deleted an invocation output from the cache"""
+        """Deletes an invocation output from the cache"""
         pass
 
     @abstractmethod
@@ -27,8 +41,22 @@ class InvocationCacheBase(ABC):
         """Clears the cache"""
         pass
 
-    @classmethod
     @abstractmethod
-    def create_key(cls, value: BaseInvocation) -> Union[int, str]:
-        """Creates the cache key for an invocation"""
+    def create_key(self, invocation: BaseInvocation) -> int:
+        """Gets the key for the invocation's cache item"""
+        pass
+
+    @abstractmethod
+    def disable(self) -> None:
+        """Disables the cache, overriding the max cache size"""
+        pass
+
+    @abstractmethod
+    def enable(self) -> None:
+        """Enables the cache, letting the the max cache size take effect"""
+        pass
+
+    @abstractmethod
+    def get_status(self) -> InvocationCacheStatus:
+        """Returns the status of the cache"""
         pass

invokeai/app/services/invocation_cache/invocation_cache_common.py

@@ -0,0 +1,9 @@
+from pydantic import BaseModel, Field
+
+
+class InvocationCacheStatus(BaseModel):
+    size: int = Field(description="The current size of the invocation cache")
+    hits: int = Field(description="The number of cache hits")
+    misses: int = Field(description="The number of cache misses")
+    enabled: bool = Field(description="Whether the invocation cache is enabled")
+    max_size: int = Field(description="The maximum size of the invocation cache")

invokeai/app/services/invocation_cache/invocation_cache_memory.py

@@ -1,70 +1,126 @@
-from queue import Queue
+from collections import OrderedDict
+from dataclasses import dataclass, field
+from threading import Lock
 from typing import Optional, Union
 
-
 from invokeai.app.invocations.baseinvocation import BaseInvocation, BaseInvocationOutput
 from invokeai.app.services.invocation_cache.invocation_cache_base import InvocationCacheBase
+from invokeai.app.services.invocation_cache.invocation_cache_common import InvocationCacheStatus
 from invokeai.app.services.invoker import Invoker
 
 
-class MemoryInvocationCache(InvocationCacheBase):
-    __cache: dict[Union[int, str], tuple[BaseInvocationOutput, str]]
-    __max_cache_size: int
-    __cache_ids: Queue
-    __invoker: Invoker
+@dataclass(order=True)
+class CachedItem:
+    invocation_output: BaseInvocationOutput = field(compare=False)
+    invocation_output_json: str = field(compare=False)
 
-    def __init__(self, max_cache_size: int = 512) -> None:
-        self.__cache = dict()
-        self.__max_cache_size = max_cache_size
-        self.__cache_ids = Queue()
+
+class MemoryInvocationCache(InvocationCacheBase):
+    _cache: OrderedDict[Union[int, str], CachedItem]
+    _max_cache_size: int
+    _disabled: bool
+    _hits: int
+    _misses: int
+    _invoker: Invoker
+    _lock: Lock
+
+    def __init__(self, max_cache_size: int = 0) -> None:
+        self._cache = OrderedDict()
+        self._max_cache_size = max_cache_size
+        self._disabled = False
+        self._hits = 0
+        self._misses = 0
+        self._lock = Lock()
 
     def start(self, invoker: Invoker) -> None:
-        self.__invoker = invoker
-        self.__invoker.services.images.on_deleted(self.delete_by_match)
+        self._invoker = invoker
+        if self._max_cache_size == 0:
+            return
+        self._invoker.services.images.on_deleted(self._delete_by_match)
+        self._invoker.services.latents.on_deleted(self._delete_by_match)
 
     def get(self, key: Union[int, str]) -> Optional[BaseInvocationOutput]:
-        if self.__max_cache_size == 0:
+        with self._lock:
+            if self._max_cache_size == 0 or self._disabled:
+                return None
+            item = self._cache.get(key, None)
+            if item is not None:
+                self._hits += 1
+                self._cache.move_to_end(key)
+                return item.invocation_output
+            self._misses += 1
             return None
 
-        item = self.__cache.get(key, None)
-        if item is not None:
-            return item[0]
+    def save(self, key: Union[int, str], invocation_output: BaseInvocationOutput) -> None:
+        with self._lock:
+            if self._max_cache_size == 0 or self._disabled or key in self._cache:
+                return
+            # If the cache is full, we need to remove the least used
+            number_to_delete = len(self._cache) + 1 - self._max_cache_size
+            self._delete_oldest_access(number_to_delete)
+            self._cache[key] = CachedItem(invocation_output, invocation_output.json())
 
-    def save(self, value: BaseInvocationOutput) -> None:
-        if self.__max_cache_size == 0:
-            return None
+    def _delete_oldest_access(self, number_to_delete: int) -> None:
+        number_to_delete = min(number_to_delete, len(self._cache))
+        for _ in range(number_to_delete):
+            self._cache.popitem(last=False)
 
-        value_json = value.json(exclude={"id"})
-        key = hash(value_json)
-
-        if key not in self.__cache:
-            self.__cache[key] = (value, value_json)
-            self.__cache_ids.put(key)
-            if self.__cache_ids.qsize() > self.__max_cache_size:
-                try:
-                    self.__cache.pop(self.__cache_ids.get())
-                except KeyError:
-                    pass
+    def _delete(self, key: Union[int, str]) -> None:
+        if self._max_cache_size == 0:
+            return
+        if key in self._cache:
+            del self._cache[key]
 
     def delete(self, key: Union[int, str]) -> None:
-        if self.__max_cache_size == 0:
-            return None
-
-        if key in self.__cache:
-            del self.__cache[key]
-
-    def delete_by_match(self, to_match: str) -> None:
-        to_delete = []
-        for name, item in self.__cache.items():
-            if to_match in item[1]:
-                to_delete.append(name)
-        for key in to_delete:
-            self.delete(key)
+        with self._lock:
+            return self._delete(key)
 
     def clear(self, *args, **kwargs) -> None:
-        self.__cache.clear()
-        self.__cache_ids = Queue()
+        with self._lock:
+            if self._max_cache_size == 0:
+                return
+            self._cache.clear()
+            self._misses = 0
+            self._hits = 0
 
-    @classmethod
-    def create_key(cls, value: BaseInvocation) -> Union[int, str]:
-        return hash(value.json(exclude={"id"}))
+    @staticmethod
+    def create_key(invocation: BaseInvocation) -> int:
+        return hash(invocation.json(exclude={"id"}))
+
+    def disable(self) -> None:
+        with self._lock:
+            if self._max_cache_size == 0:
+                return
+            self._disabled = True
+
+    def enable(self) -> None:
+        with self._lock:
+            if self._max_cache_size == 0:
+                return
+            self._disabled = False
+
+    def get_status(self) -> InvocationCacheStatus:
+        with self._lock:
+            return InvocationCacheStatus(
+                hits=self._hits,
+                misses=self._misses,
+                enabled=not self._disabled and self._max_cache_size > 0,
+                size=len(self._cache),
+                max_size=self._max_cache_size,
+            )
+
+    def _delete_by_match(self, to_match: str) -> None:
+        with self._lock:
+            if self._max_cache_size == 0:
+                return
+            keys_to_delete = set()
+            for key, cached_item in self._cache.items():
+                if to_match in cached_item.invocation_output_json:
+                    keys_to_delete.add(key)
+            if not keys_to_delete:
+                return
+            for key in keys_to_delete:
+                self._delete(key)
+            self._invoker.services.logger.debug(
+                f"Deleted {len(keys_to_delete)} cached invocation outputs for {to_match}"
+            )

invokeai/app/services/invocation_queue.py

@@ -12,9 +12,12 @@ class InvocationQueueItem(BaseModel):
     graph_execution_state_id: str = Field(description="The ID of the graph execution state")
     invocation_id: str = Field(description="The ID of the node being invoked")
     session_queue_id: str = Field(description="The ID of the session queue from which this invocation queue item came")
-    session_queue_item_id: str = Field(
+    session_queue_item_id: int = Field(
         description="The ID of session queue item from which this invocation queue item came"
     )
+    session_queue_batch_id: str = Field(
+        description="The ID of the session batch from which this invocation queue item came"
+    )
     invoke_all: bool = Field(default=False)
     timestamp: float = Field(default_factory=time.time)
 

invokeai/app/services/invoker.py

@@ -18,7 +18,12 @@ class Invoker:
         self._start()
 
     def invoke(
-        self, queue_id: str, queue_item_id: str, graph_execution_state: GraphExecutionState, invoke_all: bool = False
+        self,
+        session_queue_id: str,
+        session_queue_item_id: int,
+        session_queue_batch_id: str,
+        graph_execution_state: GraphExecutionState,
+        invoke_all: bool = False,
     ) -> Optional[str]:
         """Determines the next node to invoke and enqueues it, preparing if needed.
         Returns the id of the queued node, or `None` if there are no nodes left to enqueue."""
@@ -34,8 +39,9 @@ class Invoker:
         # Queue the invocation
         self.services.queue.put(
             InvocationQueueItem(
-                session_queue_item_id=queue_item_id,
-                session_queue_id=queue_id,
+                session_queue_id=session_queue_id,
+                session_queue_item_id=session_queue_item_id,
+                session_queue_batch_id=session_queue_batch_id,
                 graph_execution_state_id=graph_execution_state.id,
                 invocation_id=invocation.id,
                 invoke_all=invoke_all,

invokeai/app/services/latent_storage.py

@@ -3,7 +3,7 @@
 from abc import ABC, abstractmethod
 from pathlib import Path
 from queue import Queue
-from typing import Dict, Optional, Union
+from typing import Callable, Dict, Optional, Union
 
 import torch
 
@@ -11,6 +11,13 @@ import torch
 class LatentsStorageBase(ABC):
     """Responsible for storing and retrieving latents."""
 
+    _on_changed_callbacks: list[Callable[[torch.Tensor], None]]
+    _on_deleted_callbacks: list[Callable[[str], None]]
+
+    def __init__(self) -> None:
+        self._on_changed_callbacks = list()
+        self._on_deleted_callbacks = list()
+
     @abstractmethod
     def get(self, name: str) -> torch.Tensor:
         pass
@@ -23,6 +30,22 @@ class LatentsStorageBase(ABC):
     def delete(self, name: str) -> None:
         pass
 
+    def on_changed(self, on_changed: Callable[[torch.Tensor], None]) -> None:
+        """Register a callback for when an item is changed"""
+        self._on_changed_callbacks.append(on_changed)
+
+    def on_deleted(self, on_deleted: Callable[[str], None]) -> None:
+        """Register a callback for when an item is deleted"""
+        self._on_deleted_callbacks.append(on_deleted)
+
+    def _on_changed(self, item: torch.Tensor) -> None:
+        for callback in self._on_changed_callbacks:
+            callback(item)
+
+    def _on_deleted(self, item_id: str) -> None:
+        for callback in self._on_deleted_callbacks:
+            callback(item_id)
+
 
 class ForwardCacheLatentsStorage(LatentsStorageBase):
     """Caches the latest N latents in memory, writing-thorugh to and reading from underlying storage"""
@@ -33,6 +56,7 @@ class ForwardCacheLatentsStorage(LatentsStorageBase):
     __underlying_storage: LatentsStorageBase
 
     def __init__(self, underlying_storage: LatentsStorageBase, max_cache_size: int = 20):
+        super().__init__()
         self.__underlying_storage = underlying_storage
         self.__cache = dict()
         self.__cache_ids = Queue()
@@ -50,11 +74,13 @@ class ForwardCacheLatentsStorage(LatentsStorageBase):
     def save(self, name: str, data: torch.Tensor) -> None:
         self.__underlying_storage.save(name, data)
         self.__set_cache(name, data)
+        self._on_changed(data)
 
     def delete(self, name: str) -> None:
         self.__underlying_storage.delete(name)
         if name in self.__cache:
             del self.__cache[name]
+        self._on_deleted(name)
 
     def __get_cache(self, name: str) -> Optional[torch.Tensor]:
         return None if name not in self.__cache else self.__cache[name]

invokeai/app/services/model_manager_service.py

@@ -539,6 +539,7 @@ class ModelManagerService(ModelManagerServiceBase):
             context.services.events.emit_model_load_completed(
                 queue_id=context.queue_id,
                 queue_item_id=context.queue_item_id,
+                queue_batch_id=context.queue_batch_id,
                 graph_execution_state_id=context.graph_execution_state_id,
                 model_name=model_name,
                 base_model=base_model,
@@ -550,6 +551,7 @@ class ModelManagerService(ModelManagerServiceBase):
             context.services.events.emit_model_load_started(
                 queue_id=context.queue_id,
                 queue_item_id=context.queue_item_id,
+                queue_batch_id=context.queue_batch_id,
                 graph_execution_state_id=context.graph_execution_state_id,
                 model_name=model_name,
                 base_model=base_model,

invokeai/app/services/processor.py

@@ -57,6 +57,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                 except Exception as e:
                     self.__invoker.services.logger.error("Exception while retrieving session:\n%s" % e)
                     self.__invoker.services.events.emit_session_retrieval_error(
+                        queue_batch_id=queue_item.session_queue_batch_id,
                         queue_item_id=queue_item.session_queue_item_id,
                         queue_id=queue_item.session_queue_id,
                         graph_execution_state_id=queue_item.graph_execution_state_id,
@@ -70,6 +71,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                 except Exception as e:
                     self.__invoker.services.logger.error("Exception while retrieving invocation:\n%s" % e)
                     self.__invoker.services.events.emit_invocation_retrieval_error(
+                        queue_batch_id=queue_item.session_queue_batch_id,
                         queue_item_id=queue_item.session_queue_item_id,
                         queue_id=queue_item.session_queue_id,
                         graph_execution_state_id=queue_item.graph_execution_state_id,
@@ -84,6 +86,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
 
                 # Send starting event
                 self.__invoker.services.events.emit_invocation_started(
+                    queue_batch_id=queue_item.session_queue_batch_id,
                     queue_item_id=queue_item.session_queue_item_id,
                     queue_id=queue_item.session_queue_id,
                     graph_execution_state_id=graph_execution_state.id,
@@ -106,6 +109,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                                 graph_execution_state_id=graph_execution_state.id,
                                 queue_item_id=queue_item.session_queue_item_id,
                                 queue_id=queue_item.session_queue_id,
+                                queue_batch_id=queue_item.session_queue_batch_id,
                             )
                         )
 
@@ -121,6 +125,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
 
                         # Send complete event
                         self.__invoker.services.events.emit_invocation_complete(
+                            queue_batch_id=queue_item.session_queue_batch_id,
                             queue_item_id=queue_item.session_queue_item_id,
                             queue_id=queue_item.session_queue_id,
                             graph_execution_state_id=graph_execution_state.id,
@@ -150,6 +155,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                     self.__invoker.services.logger.error("Error while invoking:\n%s" % e)
                     # Send error event
                     self.__invoker.services.events.emit_invocation_error(
+                        queue_batch_id=queue_item.session_queue_batch_id,
                         queue_item_id=queue_item.session_queue_item_id,
                         queue_id=queue_item.session_queue_id,
                         graph_execution_state_id=graph_execution_state.id,
@@ -170,14 +176,16 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                 if queue_item.invoke_all and not is_complete:
                     try:
                         self.__invoker.invoke(
-                            queue_item_id=queue_item.session_queue_item_id,
-                            queue_id=queue_item.session_queue_id,
+                            session_queue_batch_id=queue_item.session_queue_batch_id,
+                            session_queue_item_id=queue_item.session_queue_item_id,
+                            session_queue_id=queue_item.session_queue_id,
                             graph_execution_state=graph_execution_state,
                             invoke_all=True,
                         )
                     except Exception as e:
                         self.__invoker.services.logger.error("Error while invoking:\n%s" % e)
                         self.__invoker.services.events.emit_invocation_error(
+                            queue_batch_id=queue_item.session_queue_batch_id,
                             queue_item_id=queue_item.session_queue_item_id,
                             queue_id=queue_item.session_queue_id,
                             graph_execution_state_id=graph_execution_state.id,
@@ -188,6 +196,7 @@ class DefaultInvocationProcessor(InvocationProcessorABC):
                         )
                 elif is_complete:
                     self.__invoker.services.events.emit_graph_execution_complete(
+                        queue_batch_id=queue_item.session_queue_batch_id,
                         queue_item_id=queue_item.session_queue_item_id,
                         queue_id=queue_item.session_queue_id,
                         graph_execution_state_id=graph_execution_state.id,

invokeai/app/services/session_processor/session_processor_default.py

@@ -1,3 +1,4 @@
+import traceback
 from threading import BoundedSemaphore
 from threading import Event as ThreadEvent
 from threading import Thread
@@ -47,20 +48,27 @@ class DefaultSessionProcessor(SessionProcessorBase):
     async def _on_queue_event(self, event: FastAPIEvent) -> None:
         event_name = event[1]["event"]
 
-        match event_name:
-            case "graph_execution_state_complete" | "invocation_error" | "session_retrieval_error" | "invocation_retrieval_error":
-                self.__queue_item = None
-                self._poll_now()
-            case "session_canceled" if self.__queue_item is not None and self.__queue_item.session_id == event[1][
-                "data"
-            ]["graph_execution_state_id"]:
-                self.__queue_item = None
-                self._poll_now()
-            case "batch_enqueued":
-                self._poll_now()
-            case "queue_cleared":
-                self.__queue_item = None
-                self._poll_now()
+        # This was a match statement, but match is not supported on python 3.9
+        if event_name in [
+            "graph_execution_state_complete",
+            "invocation_error",
+            "session_retrieval_error",
+            "invocation_retrieval_error",
+        ]:
+            self.__queue_item = None
+            self._poll_now()
+        elif (
+            event_name == "session_canceled"
+            and self.__queue_item is not None
+            and self.__queue_item.session_id == event[1]["data"]["graph_execution_state_id"]
+        ):
+            self.__queue_item = None
+            self._poll_now()
+        elif event_name == "batch_enqueued":
+            self._poll_now()
+        elif event_name == "queue_cleared":
+            self.__queue_item = None
+            self._poll_now()
 
     def resume(self) -> SessionProcessorStatus:
         if not self.__resume_event.is_set():
@@ -92,29 +100,38 @@ class DefaultSessionProcessor(SessionProcessorBase):
             self.__invoker.services.logger
             while not stop_event.is_set():
                 poll_now_event.clear()
+                try:
+                    # do not dequeue if there is already a session running
+                    if self.__queue_item is None and resume_event.is_set():
+                        queue_item = self.__invoker.services.session_queue.dequeue()
 
-                # do not dequeue if there is already a session running
-                if self.__queue_item is None and resume_event.is_set():
-                    queue_item = self.__invoker.services.session_queue.dequeue()
+                        if queue_item is not None:
+                            self.__invoker.services.logger.debug(f"Executing queue item {queue_item.item_id}")
+                            self.__queue_item = queue_item
+                            self.__invoker.services.graph_execution_manager.set(queue_item.session)
+                            self.__invoker.invoke(
+                                session_queue_batch_id=queue_item.batch_id,
+                                session_queue_id=queue_item.queue_id,
+                                session_queue_item_id=queue_item.item_id,
+                                graph_execution_state=queue_item.session,
+                                invoke_all=True,
+                            )
+                            queue_item = None
 
+                    if queue_item is None:
+                        self.__invoker.services.logger.debug("Waiting for next polling interval or event")
+                        poll_now_event.wait(POLLING_INTERVAL)
+                        continue
+                except Exception as e:
+                    self.__invoker.services.logger.error(f"Error in session processor: {e}")
                     if queue_item is not None:
-                        self.__invoker.services.logger.debug(f"Executing queue item {queue_item.item_id}")
-                        self.__queue_item = queue_item
-                        self.__invoker.services.graph_execution_manager.set(queue_item.session)
-                        self.__invoker.invoke(
-                            queue_item_id=queue_item.item_id,
-                            queue_id=queue_item.queue_id,
-                            graph_execution_state=queue_item.session,
-                            invoke_all=True,
+                        self.__invoker.services.session_queue.cancel_queue_item(
+                            queue_item.item_id, error=traceback.format_exc()
                         )
-                        queue_item = None
-
-                if queue_item is None:
-                    self.__invoker.services.logger.debug("Waiting for next polling interval or event")
                     poll_now_event.wait(POLLING_INTERVAL)
                     continue
         except Exception as e:
-            self.__invoker.services.logger.error(f"Error in session processor: {e}")
+            self.__invoker.services.logger.error(f"Fatal Error in session processor: {e}")
             pass
         finally:
             stop_event.clear()

invokeai/app/services/session_queue/session_queue_base.py

@@ -80,7 +80,7 @@ class SessionQueueBase(ABC):
         pass
 
     @abstractmethod
-    def cancel_queue_item(self, item_id: str) -> SessionQueueItem:
+    def cancel_queue_item(self, item_id: int, error: Optional[str] = None) -> SessionQueueItem:
         """Cancels a session queue item"""
         pass
 
@@ -100,13 +100,13 @@ class SessionQueueBase(ABC):
         queue_id: str,
         limit: int,
         priority: int,
-        order_id: Optional[int] = None,
+        cursor: Optional[int] = None,
         status: Optional[QUEUE_ITEM_STATUS] = None,
     ) -> CursorPaginatedResults[SessionQueueItemDTO]:
         """Gets a page of session queue items"""
         pass
 
     @abstractmethod
-    def get_queue_item(self, item_id: str) -> SessionQueueItem:
+    def get_queue_item(self, item_id: int) -> SessionQueueItem:
         """Gets a session queue item by ID"""
         pass

invokeai/app/services/session_queue/session_queue_common.py

@@ -123,6 +123,11 @@ class Batch(BaseModel):
                     raise NodeNotFoundError(f"Field {batch_data.field_name} not found in node {batch_data.node_path}")
         return values
 
+    @validator("graph")
+    def validate_graph(cls, v: Graph):
+        v.validate_self()
+        return v
+
     class Config:
         schema_extra = {
             "required": [
@@ -155,23 +160,22 @@ def get_session(queue_item_dict: dict) -> GraphExecutionState:
 class SessionQueueItemWithoutGraph(BaseModel):
     """Session queue item without the full graph. Used for serialization."""
 
-    item_id: str = Field(description="The unique identifier of the session queue item")
-    order_id: int = Field(description="The auto-incrementing ID of the session queue item")
+    item_id: int = Field(description="The identifier of the session queue item")
     status: QUEUE_ITEM_STATUS = Field(default="pending", description="The status of this queue item")
     priority: int = Field(default=0, description="The priority of this queue item")
     batch_id: str = Field(description="The ID of the batch associated with this queue item")
     session_id: str = Field(
         description="The ID of the session associated with this queue item. The session doesn't exist in graph_executions until the queue item is executed."
     )
-    field_values: Optional[list[NodeFieldValue]] = Field(
-        default=None, description="The field values that were used for this queue item"
-    )
-    queue_id: str = Field(description="The id of the queue with which this item is associated")
     error: Optional[str] = Field(default=None, description="The error message if this queue item errored")
     created_at: Union[datetime.datetime, str] = Field(description="When this queue item was created")
     updated_at: Union[datetime.datetime, str] = Field(description="When this queue item was updated")
     started_at: Optional[Union[datetime.datetime, str]] = Field(description="When this queue item was started")
     completed_at: Optional[Union[datetime.datetime, str]] = Field(description="When this queue item was completed")
+    queue_id: str = Field(description="The id of the queue with which this item is associated")
+    field_values: Optional[list[NodeFieldValue]] = Field(
+        default=None, description="The field values that were used for this queue item"
+    )
 
     @classmethod
     def from_dict(cls, queue_item_dict: dict) -> "SessionQueueItemDTO":
@@ -183,7 +187,6 @@ class SessionQueueItemWithoutGraph(BaseModel):
         schema_extra = {
             "required": [
                 "item_id",
-                "order_id",
                 "status",
                 "batch_id",
                 "queue_id",
@@ -214,7 +217,6 @@ class SessionQueueItem(SessionQueueItemWithoutGraph):
         schema_extra = {
             "required": [
                 "item_id",
-                "order_id",
                 "status",
                 "batch_id",
                 "queue_id",
@@ -235,7 +237,7 @@ class SessionQueueItem(SessionQueueItemWithoutGraph):
 
 class SessionQueueStatus(BaseModel):
     queue_id: str = Field(..., description="The ID of the queue")
-    item_id: Optional[str] = Field(description="The current queue item id")
+    item_id: Optional[int] = Field(description="The current queue item id")
     batch_id: Optional[str] = Field(description="The current queue item's batch id")
     session_id: Optional[str] = Field(description="The current queue item's session id")
     pending: int = Field(..., description="Number of queue items with status 'pending'")
@@ -388,29 +390,24 @@ def calc_session_count(batch: Batch) -> int:
 class SessionQueueValueToInsert(NamedTuple):
     """A tuple of values to insert into the session_queue table"""
 
-    item_id: str  # item_id
     queue_id: str  # queue_id
     session: str  # session json
     session_id: str  # session_id
     batch_id: str  # batch_id
     field_values: Optional[str]  # field_values json
     priority: int  # priority
-    order_id: int  # order_id
 
 
 ValuesToInsert: TypeAlias = list[SessionQueueValueToInsert]
 
 
-def prepare_values_to_insert(
-    queue_id: str, batch: Batch, priority: int, max_new_queue_items: int, order_id: int
-) -> ValuesToInsert:
+def prepare_values_to_insert(queue_id: str, batch: Batch, priority: int, max_new_queue_items: int) -> ValuesToInsert:
     values_to_insert: ValuesToInsert = []
     for session, field_values in create_session_nfv_tuples(batch, max_new_queue_items):
         # sessions must have unique id
         session.id = uuid_string()
         values_to_insert.append(
             SessionQueueValueToInsert(
-                uuid_string(),  # item_id
                 queue_id,  # queue_id
                 session.json(),  # session (json)
                 session.id,  # session_id
@@ -418,10 +415,8 @@ def prepare_values_to_insert(
                 # must use pydantic_encoder bc field_values is a list of models
                 json.dumps(field_values, default=pydantic_encoder) if field_values else None,  # field_values (json)
                 priority,  # priority
-                order_id,
             )
         )
-        order_id += 1
     return values_to_insert
 
 

invokeai/app/services/session_queue/session_queue_sqlite.py

@@ -59,13 +59,14 @@ class SqliteSessionQueue(SessionQueueBase):
 
     async def _on_session_event(self, event: FastAPIEvent) -> FastAPIEvent:
         event_name = event[1]["event"]
-        match event_name:
-            case "graph_execution_state_complete":
-                await self._handle_complete_event(event)
-            case "invocation_error" | "session_retrieval_error" | "invocation_retrieval_error":
-                await self._handle_error_event(event)
-            case "session_canceled":
-                await self._handle_cancel_event(event)
+
+        # This was a match statement, but match is not supported on python 3.9
+        if event_name == "graph_execution_state_complete":
+            await self._handle_complete_event(event)
+        elif event_name in ["invocation_error", "session_retrieval_error", "invocation_retrieval_error"]:
+            await self._handle_error_event(event)
+        elif event_name == "session_canceled":
+            await self._handle_cancel_event(event)
         return event
 
     async def _handle_complete_event(self, event: FastAPIEvent) -> None:
@@ -77,7 +78,6 @@ class SqliteSessionQueue(SessionQueueBase):
             queue_item = self.get_queue_item(item_id)
             if queue_item.status not in ["completed", "failed", "canceled"]:
                 queue_item = self._set_queue_item_status(item_id=queue_item.item_id, status="completed")
-                self.__invoker.services.events.emit_queue_item_status_changed(queue_item)
         except SessionQueueItemNotFoundError:
             return
 
@@ -86,8 +86,8 @@ class SqliteSessionQueue(SessionQueueBase):
             item_id = event[1]["data"]["queue_item_id"]
             error = event[1]["data"]["error"]
             queue_item = self.get_queue_item(item_id)
+            # always set to failed if have an error, even if previously the item was marked completed or canceled
             queue_item = self._set_queue_item_status(item_id=queue_item.item_id, status="failed", error=error)
-            self.__invoker.services.events.emit_queue_item_status_changed(queue_item)
         except SessionQueueItemNotFoundError:
             return
 
@@ -95,8 +95,8 @@ class SqliteSessionQueue(SessionQueueBase):
         try:
             item_id = event[1]["data"]["queue_item_id"]
             queue_item = self.get_queue_item(item_id)
-            queue_item = self._set_queue_item_status(item_id=queue_item.item_id, status="canceled")
-            self.__invoker.services.events.emit_queue_item_status_changed(queue_item)
+            if queue_item.status not in ["completed", "failed", "canceled"]:
+                queue_item = self._set_queue_item_status(item_id=queue_item.item_id, status="canceled")
         except SessionQueueItemNotFoundError:
             return
 
@@ -107,8 +107,7 @@ class SqliteSessionQueue(SessionQueueBase):
             self.__cursor.execute(
                 """--sql
                 CREATE TABLE IF NOT EXISTS session_queue (
-                    item_id TEXT NOT NULL PRIMARY KEY, -- the unique identifier of this queue item
-                    order_id INTEGER NOT NULL, -- used for ordering, cursor pagination
+                    item_id INTEGER PRIMARY KEY AUTOINCREMENT, -- used for ordering, cursor pagination
                     batch_id TEXT NOT NULL, -- identifier of the batch this queue item belongs to
                     queue_id TEXT NOT NULL, -- identifier of the queue this queue item belongs to
                     session_id TEXT NOT NULL UNIQUE, -- duplicated data from the session column, for ease of access
@@ -133,12 +132,6 @@ class SqliteSessionQueue(SessionQueueBase):
                 """
             )
 
-            self.__cursor.execute(
-                """--sql
-                CREATE UNIQUE INDEX IF NOT EXISTS idx_session_queue_order_id ON session_queue(order_id);
-                """
-            )
-
             self.__cursor.execute(
                 """--sql
                 CREATE UNIQUE INDEX IF NOT EXISTS idx_session_queue_session_id ON session_queue(session_id);
@@ -302,21 +295,12 @@ class SqliteSessionQueue(SessionQueueBase):
             if prepend:
                 priority = self._get_highest_priority(queue_id) + 1
 
-            self.__cursor.execute(
-                """--sql
-                SELECT MAX(order_id)
-                FROM session_queue
-                """
-            )
-            max_order_id = cast(Optional[int], self.__cursor.fetchone()[0]) or 0
-
             requested_count = calc_session_count(batch)
             values_to_insert = prepare_values_to_insert(
                 queue_id=queue_id,
                 batch=batch,
                 priority=priority,
                 max_new_queue_items=max_new_queue_items,
-                order_id=max_order_id + 1,
             )
             enqueued_count = len(values_to_insert)
 
@@ -325,8 +309,8 @@ class SqliteSessionQueue(SessionQueueBase):
 
             self.__cursor.executemany(
                 """--sql
-                INSERT INTO session_queue (item_id, queue_id, session, session_id, batch_id, field_values, priority, order_id)
-                VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+                INSERT INTO session_queue (queue_id, session, session_id, batch_id, field_values, priority)
+                VALUES (?, ?, ?, ?, ?, ?)
                 """,
                 values_to_insert,
             )
@@ -356,7 +340,7 @@ class SqliteSessionQueue(SessionQueueBase):
                 WHERE status = 'pending'
                 ORDER BY
                   priority DESC,
-                  order_id ASC
+                  item_id ASC
                 LIMIT 1
                 """
             )
@@ -370,7 +354,6 @@ class SqliteSessionQueue(SessionQueueBase):
             return None
         queue_item = SessionQueueItem.from_dict(dict(result))
         queue_item = self._set_queue_item_status(item_id=queue_item.item_id, status="in_progress")
-        self.__invoker.services.events.emit_queue_item_status_changed(queue_item)
         return queue_item
 
     def get_next(self, queue_id: str) -> Optional[SessionQueueItem]:
@@ -425,7 +408,7 @@ class SqliteSessionQueue(SessionQueueBase):
         return SessionQueueItem.from_dict(dict(result))
 
     def _set_queue_item_status(
-        self, item_id: str, status: QUEUE_ITEM_STATUS, error: Optional[str] = None
+        self, item_id: int, status: QUEUE_ITEM_STATUS, error: Optional[str] = None
     ) -> SessionQueueItem:
         try:
             self.__lock.acquire()
@@ -443,7 +426,15 @@ class SqliteSessionQueue(SessionQueueBase):
             raise
         finally:
             self.__lock.release()
-        return self.get_queue_item(item_id)
+        queue_item = self.get_queue_item(item_id)
+        batch_status = self.get_batch_status(queue_id=queue_item.queue_id, batch_id=queue_item.batch_id)
+        queue_status = self.get_queue_status(queue_id=queue_item.queue_id)
+        self.__invoker.services.events.emit_queue_item_status_changed(
+            session_queue_item=queue_item,
+            batch_status=batch_status,
+            queue_status=queue_status,
+        )
+        return queue_item
 
     def is_empty(self, queue_id: str) -> IsEmptyResult:
         try:
@@ -484,7 +475,7 @@ class SqliteSessionQueue(SessionQueueBase):
             self.__lock.release()
         return IsFullResult(is_full=is_full)
 
-    def delete_queue_item(self, item_id: str) -> SessionQueueItem:
+    def delete_queue_item(self, item_id: int) -> SessionQueueItem:
         queue_item = self.get_queue_item(item_id=item_id)
         try:
             self.__lock.acquire()
@@ -570,17 +561,18 @@ class SqliteSessionQueue(SessionQueueBase):
             self.__lock.release()
         return PruneResult(deleted=count)
 
-    def cancel_queue_item(self, item_id: str) -> SessionQueueItem:
+    def cancel_queue_item(self, item_id: int, error: Optional[str] = None) -> SessionQueueItem:
         queue_item = self.get_queue_item(item_id)
         if queue_item.status not in ["canceled", "failed", "completed"]:
-            queue_item = self._set_queue_item_status(item_id=item_id, status="canceled")
+            status = "failed" if error is not None else "canceled"
+            queue_item = self._set_queue_item_status(item_id=item_id, status=status, error=error)
             self.__invoker.services.queue.cancel(queue_item.session_id)
             self.__invoker.services.events.emit_session_canceled(
                 queue_item_id=queue_item.item_id,
                 queue_id=queue_item.queue_id,
+                queue_batch_id=queue_item.batch_id,
                 graph_execution_state_id=queue_item.session_id,
             )
-            self.__invoker.services.events.emit_queue_item_status_changed(queue_item)
         return queue_item
 
     def cancel_by_batch_ids(self, queue_id: str, batch_ids: list[str]) -> CancelByBatchIDsResult:
@@ -620,9 +612,16 @@ class SqliteSessionQueue(SessionQueueBase):
                 self.__invoker.services.events.emit_session_canceled(
                     queue_item_id=current_queue_item.item_id,
                     queue_id=current_queue_item.queue_id,
+                    queue_batch_id=current_queue_item.batch_id,
                     graph_execution_state_id=current_queue_item.session_id,
                 )
-                self.__invoker.services.events.emit_queue_item_status_changed(current_queue_item)
+                batch_status = self.get_batch_status(queue_id=queue_id, batch_id=current_queue_item.batch_id)
+                queue_status = self.get_queue_status(queue_id=queue_id)
+                self.__invoker.services.events.emit_queue_item_status_changed(
+                    session_queue_item=current_queue_item,
+                    batch_status=batch_status,
+                    queue_status=queue_status,
+                )
         except Exception:
             self.__conn.rollback()
             raise
@@ -665,9 +664,16 @@ class SqliteSessionQueue(SessionQueueBase):
                 self.__invoker.services.events.emit_session_canceled(
                     queue_item_id=current_queue_item.item_id,
                     queue_id=current_queue_item.queue_id,
+                    queue_batch_id=current_queue_item.batch_id,
                     graph_execution_state_id=current_queue_item.session_id,
                 )
-                self.__invoker.services.events.emit_queue_item_status_changed(current_queue_item)
+                batch_status = self.get_batch_status(queue_id=queue_id, batch_id=current_queue_item.batch_id)
+                queue_status = self.get_queue_status(queue_id=queue_id)
+                self.__invoker.services.events.emit_queue_item_status_changed(
+                    session_queue_item=current_queue_item,
+                    batch_status=batch_status,
+                    queue_status=queue_status,
+                )
         except Exception:
             self.__conn.rollback()
             raise
@@ -675,7 +681,7 @@ class SqliteSessionQueue(SessionQueueBase):
             self.__lock.release()
         return CancelByQueueIDResult(canceled=count)
 
-    def get_queue_item(self, item_id: str) -> SessionQueueItem:
+    def get_queue_item(self, item_id: int) -> SessionQueueItem:
         try:
             self.__lock.acquire()
             self.__cursor.execute(
@@ -701,14 +707,14 @@ class SqliteSessionQueue(SessionQueueBase):
         queue_id: str,
         limit: int,
         priority: int,
-        order_id: Optional[int] = None,
+        cursor: Optional[int] = None,
         status: Optional[QUEUE_ITEM_STATUS] = None,
     ) -> CursorPaginatedResults[SessionQueueItemDTO]:
         try:
+            item_id = cursor
             self.__lock.acquire()
             query = """--sql
                 SELECT item_id,
-                    order_id,
                     status,
                     priority,
                     field_values,
@@ -731,16 +737,16 @@ class SqliteSessionQueue(SessionQueueBase):
                     """
                 params.append(status)
 
-            if order_id is not None:
+            if item_id is not None:
                 query += """--sql
-                    AND (priority < ?) OR (priority = ? AND order_id > ?)
+                    AND (priority < ?) OR (priority = ? AND item_id > ?)
                     """
-                params.extend([priority, priority, order_id])
+                params.extend([priority, priority, item_id])
 
             query += """--sql
                 ORDER BY
                   priority DESC,
-                  order_id ASC
+                  item_id ASC
                 LIMIT ?
                 """
             params.append(limit + 1)

invokeai/app/util/controlnet_utils.py

@@ -265,22 +265,41 @@ def np_img_resize(np_img: np.ndarray, resize_mode: str, h: int, w: int, device:
 
 
 def prepare_control_image(
-    # image used to be Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor, List[torch.Tensor]]
-    # but now should be able to assume that image is a single PIL.Image, which simplifies things
     image: Image,
-    # FIXME: need to fix hardwiring of width and height, change to basing on latents dimensions?
-    # latents_to_match_resolution, # TorchTensor of shape (batch_size, 3, height, width)
-    width=512,  # should be 8 * latent.shape[3]
-    height=512,  # should be 8 * latent height[2]
-    # batch_size=1, # currently no batching
-    # num_images_per_prompt=1, # currently only single image
+    width: int,
+    height: int,
+    num_channels: int = 3,
     device="cuda",
     dtype=torch.float16,
     do_classifier_free_guidance=True,
     control_mode="balanced",
     resize_mode="just_resize_simple",
 ):
-    # FIXME: implement "crop_resize_simple" and "fill_resize_simple", or pull them out
+    """Pre-process images for ControlNets or T2I-Adapters.
+
+    Args:
+        image (Image): The PIL image to pre-process.
+        width (int): The target width in pixels.
+        height (int): The target height in pixels.
+        num_channels (int, optional): The target number of image channels. This is achieved by converting the input
+            image to RGB, then naively taking the first `num_channels` channels. The primary use case is converting a
+            RGB image to a single-channel grayscale image. Raises if `num_channels` cannot be achieved. Defaults to 3.
+        device (str, optional): The target device for the output image. Defaults to "cuda".
+        dtype (_type_, optional): The dtype for the output image. Defaults to torch.float16.
+        do_classifier_free_guidance (bool, optional): If True, repeat the output image along the batch dimension.
+            Defaults to True.
+        control_mode (str, optional): Defaults to "balanced".
+        resize_mode (str, optional): Defaults to "just_resize_simple".
+
+    Raises:
+        NotImplementedError: If resize_mode == "crop_resize_simple".
+        NotImplementedError: If resize_mode == "fill_resize_simple".
+        ValueError: If `resize_mode` is not recognized.
+        ValueError: If `num_channels` is out of range.
+
+    Returns:
+        torch.Tensor: The pre-processed input tensor.
+    """
     if (
         resize_mode == "just_resize_simple"
         or resize_mode == "crop_resize_simple"
@@ -289,10 +308,10 @@ def prepare_control_image(
         image = image.convert("RGB")
         if resize_mode == "just_resize_simple":
             image = image.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])
-        elif resize_mode == "crop_resize_simple":  # not yet implemented
-            pass
-        elif resize_mode == "fill_resize_simple":  # not yet implemented
-            pass
+        elif resize_mode == "crop_resize_simple":
+            raise NotImplementedError(f"prepare_control_image is not implemented for resize_mode='{resize_mode}'.")
+        elif resize_mode == "fill_resize_simple":
+            raise NotImplementedError(f"prepare_control_image is not implemented for resize_mode='{resize_mode}'.")
         nimage = np.array(image)
         nimage = nimage[None, :]
         nimage = np.concatenate([nimage], axis=0)
@@ -313,9 +332,11 @@ def prepare_control_image(
             device=device,
         )
     else:
-        pass
-        print("ERROR: invalid resize_mode ==> ", resize_mode)
-        exit(1)
+        raise ValueError(f"Unsupported resize_mode: '{resize_mode}'.")
+
+    if timage.shape[1] < num_channels or num_channels <= 0:
+        raise ValueError(f"Cannot achieve the target of num_channels={num_channels}.")
+    timage = timage[:, :num_channels, :, :]
 
     timage = timage.to(device=device, dtype=dtype)
     cfg_injection = control_mode == "more_control" or control_mode == "unbalanced"

invokeai/app/util/step_callback.py

@@ -112,6 +112,7 @@ def stable_diffusion_step_callback(
     context.services.events.emit_generator_progress(
         queue_id=context.queue_id,
         queue_item_id=context.queue_item_id,
+        queue_batch_id=context.queue_batch_id,
         graph_execution_state_id=context.graph_execution_state_id,
         node=node,
         source_node_id=source_node_id,

invokeai/assets/fonts/inter/LICENSE.txt

@@ -0,0 +1,94 @@
+Copyright (c) 2016-2020 The Inter Project Authors.
+"Inter" is trademark of Rasmus Andersson.
+https://github.com/rsms/inter
+
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+http://scripts.sil.org/OFL
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION AND CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

invokeai/backend/image_util/invoke_metadata.py

@@ -0,0 +1,46 @@
+# Copyright (c) 2023 Lincoln D. Stein and the InvokeAI Development Team
+
+"""Very simple functions to fetch and print metadata from InvokeAI-generated images."""
+
+import json
+import sys
+from pathlib import Path
+from typing import Any, Dict
+
+from PIL import Image
+
+
+def get_invokeai_metadata(image_path: Path) -> Dict[str, Any]:
+    """
+    Retrieve "invokeai_metadata" field from png image.
+
+    :param image_path: Path to the image to read metadata from.
+    May raise:
+      OSError -- image path not found
+      KeyError -- image doesn't contain the metadata field
+    """
+    image: Image = Image.open(image_path)
+    return json.loads(image.text["invokeai_metadata"])
+
+
+def print_invokeai_metadata(image_path: Path):
+    """Pretty-print the metadata."""
+    try:
+        metadata = get_invokeai_metadata(image_path)
+        print(f"{image_path}:\n{json.dumps(metadata, sort_keys=True, indent=4)}")
+    except OSError:
+        print(f"{image_path}:\nNo file found.")
+    except KeyError:
+        print(f"{image_path}:\nNo metadata found.")
+    print()
+
+
+def main():
+    """Run the command-line utility."""
+    image_paths = sys.argv[1:]
+    if not image_paths:
+        print(f"Usage: {Path(sys.argv[0]).name} image1 image2 image3 ...")
+        print("\nPretty-print InvokeAI image metadata from the listed png files.")
+        sys.exit(-1)
+    for img in image_paths:
+        print_invokeai_metadata(img)

invokeai/backend/install/invokeai_configure.py

@@ -70,7 +70,6 @@ def get_literal_fields(field) -> list[Any]:
 config = InvokeAIAppConfig.get_config()
 
 Model_dir = "models"
-
 Default_config_file = config.model_conf_path
 SD_Configs = config.legacy_conf_path
 
@@ -93,7 +92,7 @@ INIT_FILE_PREAMBLE = """# InvokeAI initialization file
 # or renaming it and then running invokeai-configure again.
 """
 
-logger = InvokeAILogger.getLogger()
+logger = InvokeAILogger.get_logger()
 
 
 class DummyWidgetValue(Enum):
@@ -458,7 +457,7 @@ Use cursor arrows to make a checkbox selection, and space to toggle.
         )
         self.add_widget_intelligent(
             npyscreen.TitleFixedText,
-            name="Model RAM cache size (GB). Make this at least large enough to hold a single full model.",
+            name="Model RAM cache size (GB). Make this at least large enough to hold a single full model (2GB for SD-1, 6GB for SDXL).",
             begin_entry_at=0,
             editable=False,
             color="CONTROL",
@@ -651,8 +650,19 @@ def edit_opts(program_opts: Namespace, invokeai_opts: Namespace) -> argparse.Nam
     return editApp.new_opts()
 
 
+def default_ramcache() -> float:
+    """Run a heuristic for the default RAM cache based on installed RAM."""
+
+    # Note that on my 64 GB machine, psutil.virtual_memory().total gives 62 GB,
+    # So we adjust everthing down a bit.
+    return (
+        15.0 if MAX_RAM >= 60 else 7.5 if MAX_RAM >= 30 else 4 if MAX_RAM >= 14 else 2.1
+    )  # 2.1 is just large enough for sd 1.5 ;-)
+
+
 def default_startup_options(init_file: Path) -> Namespace:
     opts = InvokeAIAppConfig.get_config()
+    opts.ram = opts.ram or default_ramcache()
     return opts
 
 
@@ -894,7 +904,7 @@ def main():
     if opt.full_precision:
         invoke_args.extend(["--precision", "float32"])
     config.parse_args(invoke_args)
-    logger = InvokeAILogger().getLogger(config=config)
+    logger = InvokeAILogger().get_logger(config=config)
 
     errors = set()
 

invokeai/backend/install/model_install_backend.py

@@ -2,6 +2,7 @@
 Utility (backend) functions used by model_install.py
 """
 import os
+import re
 import shutil
 import warnings
 from dataclasses import dataclass, field
@@ -30,7 +31,7 @@ warnings.filterwarnings("ignore")
 
 # --------------------------globals-----------------------
 config = InvokeAIAppConfig.get_config()
-logger = InvokeAILogger.getLogger(name="InvokeAI")
+logger = InvokeAILogger.get_logger(name="InvokeAI")
 
 # the initial "configs" dir is now bundled in the `invokeai.configs` package
 Dataset_path = Path(configs.__path__[0]) / "INITIAL_MODELS.yaml"
@@ -47,8 +48,14 @@ Config_preamble = """
 
 LEGACY_CONFIGS = {
     BaseModelType.StableDiffusion1: {
-        ModelVariantType.Normal: "v1-inference.yaml",
-        ModelVariantType.Inpaint: "v1-inpainting-inference.yaml",
+        ModelVariantType.Normal: {
+            SchedulerPredictionType.Epsilon: "v1-inference.yaml",
+            SchedulerPredictionType.VPrediction: "v1-inference-v.yaml",
+        },
+        ModelVariantType.Inpaint: {
+            SchedulerPredictionType.Epsilon: "v1-inpainting-inference.yaml",
+            SchedulerPredictionType.VPrediction: "v1-inpainting-inference-v.yaml",
+        },
     },
     BaseModelType.StableDiffusion2: {
         ModelVariantType.Normal: {
@@ -69,14 +76,6 @@ LEGACY_CONFIGS = {
 }
 
 
-@dataclass
-class ModelInstallList:
-    """Class for listing models to be installed/removed"""
-
-    install_models: List[str] = field(default_factory=list)
-    remove_models: List[str] = field(default_factory=list)
-
-
 @dataclass
 class InstallSelections:
     install_models: List[str] = field(default_factory=list)
@@ -90,10 +89,12 @@ class ModelLoadInfo:
     base_type: BaseModelType
     path: Optional[Path] = None
     repo_id: Optional[str] = None
+    subfolder: Optional[str] = None
     description: str = ""
     installed: bool = False
     recommended: bool = False
     default: bool = False
+    requires: Optional[List[str]] = field(default_factory=list)
 
 
 class ModelInstall(object):
@@ -127,12 +128,13 @@ class ModelInstall(object):
             value["name"] = name
             value["base_type"] = base
             value["model_type"] = model_type
-            model_dict[key] = ModelLoadInfo(**value)
+            model_info = ModelLoadInfo(**value)
+            if model_info.subfolder and model_info.repo_id:
+                model_info.repo_id += f":{model_info.subfolder}"
+            model_dict[key] = model_info
 
         # supplement with entries in models.yaml
         installed_models = [x for x in self.mgr.list_models()]
-        # suppresses autoloaded models
-        # installed_models = [x for x in self.mgr.list_models() if not self._is_autoloaded(x)]
 
         for md in installed_models:
             base = md["base_model"]
@@ -164,9 +166,12 @@ class ModelInstall(object):
 
     def list_models(self, model_type):
         installed = self.mgr.list_models(model_type=model_type)
+        print()
         print(f"Installed models of type `{model_type}`:")
+        print(f"{'Model Key':50} Model Path")
         for i in installed:
-            print(f"{i['model_name']}\t{i['base_model']}\t{i['path']}")
+            print(f"{'/'.join([i['base_model'],i['model_type'],i['model_name']]):50} {i['path']}")
+        print()
 
     # logic here a little reversed to maintain backward compatibility
     def starter_models(self, all_models: bool = False) -> Set[str]:
@@ -204,6 +209,8 @@ class ModelInstall(object):
             job += 1
 
         # add requested models
+        self._remove_installed(selections.install_models)
+        self._add_required_models(selections.install_models)
         for path in selections.install_models:
             logger.info(f"Installing {path} [{job}/{jobs}]")
             try:
@@ -263,6 +270,26 @@ class ModelInstall(object):
 
         return models_installed
 
+    def _remove_installed(self, model_list: List[str]):
+        all_models = self.all_models()
+        for path in model_list:
+            key = self.reverse_paths.get(path)
+            if key and all_models[key].installed:
+                logger.warning(f"{path} already installed. Skipping.")
+                model_list.remove(path)
+
+    def _add_required_models(self, model_list: List[str]):
+        additional_models = []
+        all_models = self.all_models()
+        for path in model_list:
+            if not (key := self.reverse_paths.get(path)):
+                continue
+            for requirement in all_models[key].requires:
+                requirement_key = self.reverse_paths.get(requirement)
+                if not all_models[requirement_key].installed:
+                    additional_models.append(requirement)
+        model_list.extend(additional_models)
+
     # install a model from a local path. The optional info parameter is there to prevent
     # the model from being probed twice in the event that it has already been probed.
     def _install_path(self, path: Path, info: ModelProbeInfo = None) -> AddModelResult:
@@ -286,7 +313,7 @@ class ModelInstall(object):
             location = download_with_resume(url, Path(staging))
             if not location:
                 logger.error(f"Unable to download {url}. Skipping.")
-            info = ModelProbe().heuristic_probe(location)
+            info = ModelProbe().heuristic_probe(location, self.prediction_helper)
             dest = self.config.models_path / info.base_type.value / info.model_type.value / location.name
             dest.parent.mkdir(parents=True, exist_ok=True)
             models_path = shutil.move(location, dest)
@@ -295,36 +322,63 @@ class ModelInstall(object):
         return self._install_path(Path(models_path), info)
 
     def _install_repo(self, repo_id: str) -> AddModelResult:
+        # hack to recover models stored in subfolders --
+        # Required to get the "v2" model of monster-labs/control_v1p_sd15_qrcode_monster
+        subfolder = None
+        if match := re.match(r"^([^/]+/[^/]+):(\w+)$", repo_id):
+            repo_id = match.group(1)
+            subfolder = match.group(2)
+
         hinfo = HfApi().model_info(repo_id)
 
         # we try to figure out how to download this most economically
         # list all the files in the repo
         files = [x.rfilename for x in hinfo.siblings]
+        if subfolder:
+            files = [x for x in files if x.startswith(f"{subfolder}/")]
+        prefix = f"{subfolder}/" if subfolder else ""
+
         location = None
 
         with TemporaryDirectory(dir=self.config.models_path) as staging:
             staging = Path(staging)
-            if "model_index.json" in files:
-                location = self._download_hf_pipeline(repo_id, staging)  # pipeline
-            elif "unet/model.onnx" in files:
+            if f"{prefix}model_index.json" in files:
+                location = self._download_hf_pipeline(repo_id, staging, subfolder=subfolder)  # pipeline
+            elif f"{prefix}unet/model.onnx" in files:
                 location = self._download_hf_model(repo_id, files, staging)
             else:
                 for suffix in ["safetensors", "bin"]:
-                    if f"pytorch_lora_weights.{suffix}" in files:
-                        location = self._download_hf_model(repo_id, ["pytorch_lora_weights.bin"], staging)  # LoRA
+                    if f"{prefix}pytorch_lora_weights.{suffix}" in files:
+                        location = self._download_hf_model(
+                            repo_id, ["pytorch_lora_weights.bin"], staging, subfolder=subfolder
+                        )  # LoRA
                         break
                     elif (
-                        self.config.precision == "float16" and f"diffusion_pytorch_model.fp16.{suffix}" in files
+                        self.config.precision == "float16" and f"{prefix}diffusion_pytorch_model.fp16.{suffix}" in files
                     ):  # vae, controlnet or some other standalone
                         files = ["config.json", f"diffusion_pytorch_model.fp16.{suffix}"]
-                        location = self._download_hf_model(repo_id, files, staging)
+                        location = self._download_hf_model(repo_id, files, staging, subfolder=subfolder)
                         break
-                    elif f"diffusion_pytorch_model.{suffix}" in files:
+                    elif f"{prefix}diffusion_pytorch_model.{suffix}" in files:
                         files = ["config.json", f"diffusion_pytorch_model.{suffix}"]
-                        location = self._download_hf_model(repo_id, files, staging)
+                        location = self._download_hf_model(repo_id, files, staging, subfolder=subfolder)
                         break
-                    elif f"learned_embeds.{suffix}" in files:
-                        location = self._download_hf_model(repo_id, [f"learned_embeds.{suffix}"], staging)
+                    elif f"{prefix}learned_embeds.{suffix}" in files:
+                        location = self._download_hf_model(
+                            repo_id, [f"learned_embeds.{suffix}"], staging, subfolder=subfolder
+                        )
+                        break
+                    elif (
+                        f"{prefix}image_encoder.txt" in files and f"{prefix}ip_adapter.{suffix}" in files
+                    ):  # IP-Adapter
+                        files = ["image_encoder.txt", f"ip_adapter.{suffix}"]
+                        location = self._download_hf_model(repo_id, files, staging, subfolder=subfolder)
+                        break
+                    elif f"{prefix}model.{suffix}" in files and f"{prefix}config.json" in files:
+                        # This elif-condition is pretty fragile, but it is intended to handle CLIP Vision models hosted
+                        # by InvokeAI for use with IP-Adapters.
+                        files = ["config.json", f"model.{suffix}"]
+                        location = self._download_hf_model(repo_id, files, staging, subfolder=subfolder)
                         break
             if not location:
                 logger.warning(f"Could not determine type of repo {repo_id}. Skipping install.")
@@ -383,7 +437,7 @@ class ModelInstall(object):
                     possible_conf = path.with_suffix(".yaml")
                     if possible_conf.exists():
                         legacy_conf = str(self.relative_to_root(possible_conf))
-                    elif info.base_type == BaseModelType.StableDiffusion2:
+                    elif info.base_type in [BaseModelType.StableDiffusion1, BaseModelType.StableDiffusion2]:
                         legacy_conf = Path(
                             self.config.legacy_conf_dir,
                             LEGACY_CONFIGS[info.base_type][info.variant_type][info.prediction_type],
@@ -411,9 +465,9 @@ class ModelInstall(object):
         else:
             return path
 
-    def _download_hf_pipeline(self, repo_id: str, staging: Path) -> Path:
+    def _download_hf_pipeline(self, repo_id: str, staging: Path, subfolder: str = None) -> Path:
         """
-        This retrieves a StableDiffusion model from cache or remote and then
+        Retrieve a StableDiffusion model from cache or remote and then
         does a save_pretrained() to the indicated staging area.
         """
         _, name = repo_id.split("/")
@@ -428,6 +482,7 @@ class ModelInstall(object):
                     variant=variant,
                     torch_dtype=precision,
                     safety_checker=None,
+                    subfolder=subfolder,
                 )
             except Exception as e:  # most errors are due to fp16 not being present. Fix this to catch other errors
                 if "fp16" not in str(e):
@@ -442,7 +497,7 @@ class ModelInstall(object):
         model.save_pretrained(staging / name, safe_serialization=True)
         return staging / name
 
-    def _download_hf_model(self, repo_id: str, files: List[str], staging: Path) -> Path:
+    def _download_hf_model(self, repo_id: str, files: List[str], staging: Path, subfolder: None) -> Path:
         _, name = repo_id.split("/")
         location = staging / name
         paths = list()
@@ -453,7 +508,7 @@ class ModelInstall(object):
                 model_dir=location / filePath.parent,
                 model_name=filePath.name,
                 access_token=self.access_token,
-                subfolder=filePath.parent,
+                subfolder=filePath.parent / subfolder if subfolder else filePath.parent,
             )
             if p:
                 paths.append(p)
@@ -482,7 +537,7 @@ def yes_or_no(prompt: str, default_yes=True):
 
 # ---------------------------------------------
 def hf_download_from_pretrained(model_class: object, model_name: str, destination: Path, **kwargs):
-    logger = InvokeAILogger.getLogger("InvokeAI")
+    logger = InvokeAILogger.get_logger("InvokeAI")
     logger.addFilter(lambda x: "fp16 is not a valid" not in x.getMessage())
 
     model = model_class.from_pretrained(
@@ -534,14 +589,17 @@ def hf_download_with_resume(
         logger.info(f"{model_name}: Downloading...")
 
     try:
-        with open(model_dest, open_mode) as file, tqdm(
-            desc=model_name,
-            initial=exist_size,
-            total=total + exist_size,
-            unit="iB",
-            unit_scale=True,
-            unit_divisor=1000,
-        ) as bar:
+        with (
+            open(model_dest, open_mode) as file,
+            tqdm(
+                desc=model_name,
+                initial=exist_size,
+                total=total + exist_size,
+                unit="iB",
+                unit_scale=True,
+                unit_divisor=1000,
+            ) as bar,
+        ):
             for data in resp.iter_content(chunk_size=1024):
                 size = file.write(data)
                 bar.update(size)

invokeai/backend/ip_adapter/README.md

@@ -0,0 +1,45 @@
+# IP-Adapter Model Formats
+
+The official IP-Adapter models are released here: [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter)
+
+This official model repo does not integrate well with InvokeAI's current approach to model management, so we have defined a new file structure for IP-Adapter models. The InvokeAI format is described below.
+
+## CLIP Vision Models
+
+CLIP Vision models are organized in `diffusers`` format. The expected directory structure is:
+
+```bash
+ip_adapter_sd_image_encoder/
+├── config.json
+└── model.safetensors
+```
+
+## IP-Adapter Models
+
+IP-Adapter models are stored in a directory containing two files
+- `image_encoder.txt`: A text file containing the model identifier for the CLIP Vision encoder that is intended to be used with this IP-Adapter model.
+- `ip_adapter.bin`: The IP-Adapter weights.
+
+Sample directory structure:
+```bash
+ip_adapter_sd15/
+├── image_encoder.txt
+└── ip_adapter.bin
+```
+
+### Why save the weights in a .safetensors file?
+
+The weights in `ip_adapter.bin` are stored in a nested dict, which is not supported by `safetensors`. This could be solved by splitting `ip_adapter.bin` into multiple files, but for now we have decided to maintain consistency with the checkpoint structure used in the official [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter) repo.
+
+## InvokeAI Hosted IP-Adapters
+
+Image Encoders:
+- [InvokeAI/ip_adapter_sd_image_encoder](https://huggingface.co/InvokeAI/ip_adapter_sd_image_encoder)
+- [InvokeAI/ip_adapter_sdxl_image_encoder](https://huggingface.co/InvokeAI/ip_adapter_sdxl_image_encoder)
+
+IP-Adapters:
+- [InvokeAI/ip_adapter_sd15](https://huggingface.co/InvokeAI/ip_adapter_sd15)
+- [InvokeAI/ip_adapter_plus_sd15](https://huggingface.co/InvokeAI/ip_adapter_plus_sd15)
+- [InvokeAI/ip_adapter_plus_face_sd15](https://huggingface.co/InvokeAI/ip_adapter_plus_face_sd15)
+- [InvokeAI/ip_adapter_sdxl](https://huggingface.co/InvokeAI/ip_adapter_sdxl)
+- [InvokeAI/ip_adapter_sdxl_vit_h](https://huggingface.co/InvokeAI/ip_adapter_sdxl_vit_h)

invokeai/backend/ip_adapter/attention_processor.py

@@ -0,0 +1,165 @@
+# copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
+#   and modified as needed
+
+# tencent-ailab comment:
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.attention_processor import AttnProcessor2_0 as DiffusersAttnProcessor2_0
+
+from invokeai.backend.ip_adapter.ip_attention_weights import IPAttentionProcessorWeights
+
+
+# Create a version of AttnProcessor2_0 that is a sub-class of nn.Module. This is required for IP-Adapter state_dict
+# loading.
+class AttnProcessor2_0(DiffusersAttnProcessor2_0, nn.Module):
+    def __init__(self):
+        DiffusersAttnProcessor2_0.__init__(self)
+        nn.Module.__init__(self)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        ip_adapter_image_prompt_embeds=None,
+    ):
+        """Re-definition of DiffusersAttnProcessor2_0.__call__(...) that accepts and ignores the
+        ip_adapter_image_prompt_embeds parameter.
+        """
+        return DiffusersAttnProcessor2_0.__call__(
+            self, attn, hidden_states, encoder_hidden_states, attention_mask, temb
+        )
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+    """
+
+    def __init__(self, weights: list[IPAttentionProcessorWeights], scales: list[float]):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        assert len(weights) == len(scales)
+
+        self._weights = weights
+        self._scales = scales
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        ip_adapter_image_prompt_embeds=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if encoder_hidden_states is not None:
+            # If encoder_hidden_states is not None, then we are doing cross-attention, not self-attention. In this case,
+            # we will apply IP-Adapter conditioning. We validate the inputs for IP-Adapter conditioning here.
+            assert ip_adapter_image_prompt_embeds is not None
+            assert len(ip_adapter_image_prompt_embeds) == len(self._weights)
+
+            for ipa_embed, ipa_weights, scale in zip(ip_adapter_image_prompt_embeds, self._weights, self._scales):
+                # The batch dimensions should match.
+                assert ipa_embed.shape[0] == encoder_hidden_states.shape[0]
+                # The channel dimensions should match.
+                assert ipa_embed.shape[2] == encoder_hidden_states.shape[2]
+
+                ip_hidden_states = ipa_embed
+
+                ip_key = ipa_weights.to_k_ip(ip_hidden_states)
+                ip_value = ipa_weights.to_v_ip(ip_hidden_states)
+
+                ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+                ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+                # The output of sdpa has shape: (batch, num_heads, seq_len, head_dim)
+                # TODO: add support for attn.scale when we move to Torch 2.1
+                ip_hidden_states = F.scaled_dot_product_attention(
+                    query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+                )
+
+                ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+                ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+                hidden_states = hidden_states + scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

invokeai/backend/ip_adapter/ip_adapter.py

@@ -0,0 +1,167 @@
+# copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
+#   and modified as needed
+
+from typing import Optional, Union
+
+import torch
+from PIL import Image
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+from invokeai.backend.ip_adapter.ip_attention_weights import IPAttentionWeights
+from invokeai.backend.model_management.models.base import calc_model_size_by_data
+
+from .resampler import Resampler
+
+
+class ImageProjModel(torch.nn.Module):
+    """Image Projection Model"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    @classmethod
+    def from_state_dict(cls, state_dict: dict[torch.Tensor], clip_extra_context_tokens=4):
+        """Initialize an ImageProjModel from a state_dict.
+
+        The cross_attention_dim and clip_embeddings_dim are inferred from the shape of the tensors in the state_dict.
+
+        Args:
+            state_dict (dict[torch.Tensor]): The state_dict of model weights.
+            clip_extra_context_tokens (int, optional): Defaults to 4.
+
+        Returns:
+            ImageProjModel
+        """
+        cross_attention_dim = state_dict["norm.weight"].shape[0]
+        clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+
+        model = cls(cross_attention_dim, clip_embeddings_dim, clip_extra_context_tokens)
+
+        model.load_state_dict(state_dict)
+        return model
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class IPAdapter:
+    """IP-Adapter: https://arxiv.org/pdf/2308.06721.pdf"""
+
+    def __init__(
+        self,
+        state_dict: dict[str, torch.Tensor],
+        device: torch.device,
+        dtype: torch.dtype = torch.float16,
+        num_tokens: int = 4,
+    ):
+        self.device = device
+        self.dtype = dtype
+
+        self._num_tokens = num_tokens
+
+        self._clip_image_processor = CLIPImageProcessor()
+
+        self._image_proj_model = self._init_image_proj_model(state_dict["image_proj"])
+
+        self.attn_weights = IPAttentionWeights.from_state_dict(state_dict["ip_adapter"]).to(
+            self.device, dtype=self.dtype
+        )
+
+    def to(self, device: torch.device, dtype: Optional[torch.dtype] = None):
+        self.device = device
+        if dtype is not None:
+            self.dtype = dtype
+
+        self._image_proj_model.to(device=self.device, dtype=self.dtype)
+        self.attn_weights.to(device=self.device, dtype=self.dtype)
+
+    def calc_size(self):
+        return calc_model_size_by_data(self._image_proj_model) + calc_model_size_by_data(self.attn_weights)
+
+    def _init_image_proj_model(self, state_dict):
+        return ImageProjModel.from_state_dict(state_dict, self._num_tokens).to(self.device, dtype=self.dtype)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image, image_encoder: CLIPVisionModelWithProjection):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self._clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image_embeds = image_encoder(clip_image.to(self.device, dtype=self.dtype)).image_embeds
+        image_prompt_embeds = self._image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self._image_proj_model(torch.zeros_like(clip_image_embeds))
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class IPAdapterPlus(IPAdapter):
+    """IP-Adapter with fine-grained features"""
+
+    def _init_image_proj_model(self, state_dict):
+        return Resampler.from_state_dict(
+            state_dict=state_dict,
+            depth=4,
+            dim_head=64,
+            heads=12,
+            num_queries=self._num_tokens,
+            ff_mult=4,
+        ).to(self.device, dtype=self.dtype)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image, image_encoder: CLIPVisionModelWithProjection):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self._clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image = clip_image.to(self.device, dtype=self.dtype)
+        clip_image_embeds = image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        image_prompt_embeds = self._image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = image_encoder(torch.zeros_like(clip_image), output_hidden_states=True).hidden_states[
+            -2
+        ]
+        uncond_image_prompt_embeds = self._image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class IPAdapterPlusXL(IPAdapterPlus):
+    """IP-Adapter Plus for SDXL."""
+
+    def _init_image_proj_model(self, state_dict):
+        return Resampler.from_state_dict(
+            state_dict=state_dict,
+            depth=4,
+            dim_head=64,
+            heads=20,
+            num_queries=self._num_tokens,
+            ff_mult=4,
+        ).to(self.device, dtype=self.dtype)
+
+
+def build_ip_adapter(
+    ip_adapter_ckpt_path: str, device: torch.device, dtype: torch.dtype = torch.float16
+) -> Union[IPAdapter, IPAdapterPlus]:
+    state_dict = torch.load(ip_adapter_ckpt_path, map_location="cpu")
+
+    # Determine if the state_dict is from an IPAdapter or IPAdapterPlus based on the image_proj weights that it
+    # contains.
+    is_plus = "proj.weight" not in state_dict["image_proj"]
+
+    if is_plus:
+        cross_attention_dim = state_dict["ip_adapter"]["1.to_k_ip.weight"].shape[-1]
+        if cross_attention_dim == 768:
+            # SD1 IP-Adapter Plus
+            return IPAdapterPlus(state_dict, device=device, dtype=dtype)
+        elif cross_attention_dim == 2048:
+            # SDXL IP-Adapter Plus
+            return IPAdapterPlusXL(state_dict, device=device, dtype=dtype)
+        else:
+            raise Exception(f"Unsupported IP-Adapter Plus cross-attention dimension: {cross_attention_dim}.")
+    else:
+        return IPAdapter(state_dict, device=device, dtype=dtype)

invokeai/backend/ip_adapter/ip_attention_weights.py

@@ -0,0 +1,46 @@
+import torch
+
+
+class IPAttentionProcessorWeights(torch.nn.Module):
+    """The IP-Adapter weights for a single attention processor.
+
+    This class is a torch.nn.Module sub-class to facilitate loading from a state_dict. It does not have a forward(...)
+    method.
+    """
+
+    def __init__(self, in_dim: int, out_dim: int):
+        super().__init__()
+        self.to_k_ip = torch.nn.Linear(in_dim, out_dim, bias=False)
+        self.to_v_ip = torch.nn.Linear(in_dim, out_dim, bias=False)
+
+
+class IPAttentionWeights(torch.nn.Module):
+    """A collection of all the `IPAttentionProcessorWeights` objects for an IP-Adapter model.
+
+    This class is a torch.nn.Module sub-class so that it inherits the `.to(...)` functionality. It does not have a
+    forward(...) method.
+    """
+
+    def __init__(self, weights: torch.nn.ModuleDict):
+        super().__init__()
+        self._weights = weights
+
+    def get_attention_processor_weights(self, idx: int) -> IPAttentionProcessorWeights:
+        """Get the `IPAttentionProcessorWeights` for the idx'th attention processor."""
+        # Cast to int first, because we expect the key to represent an int. Then cast back to str, because
+        # `torch.nn.ModuleDict` only supports str keys.
+        return self._weights[str(int(idx))]
+
+    @classmethod
+    def from_state_dict(cls, state_dict: dict[str, torch.Tensor]):
+        attn_proc_weights: dict[str, IPAttentionProcessorWeights] = {}
+
+        for tensor_name, tensor in state_dict.items():
+            if "to_k_ip.weight" in tensor_name:
+                index = str(int(tensor_name.split(".")[0]))
+                attn_proc_weights[index] = IPAttentionProcessorWeights(tensor.shape[1], tensor.shape[0])
+
+        attn_proc_weights_module = torch.nn.ModuleDict(attn_proc_weights)
+        attn_proc_weights_module.load_state_dict(state_dict)
+
+        return cls(attn_proc_weights_module)

invokeai/backend/ip_adapter/resampler.py

@@ -0,0 +1,158 @@
+# copied from https://github.com/tencent-ailab/IP-Adapter (Apache License 2.0)
+
+# tencent ailab comment: modified from
+# https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+import math
+
+import torch
+import torch.nn as nn
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+    ):
+        super().__init__()
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    @classmethod
+    def from_state_dict(cls, state_dict: dict[torch.Tensor], depth=8, dim_head=64, heads=16, num_queries=8, ff_mult=4):
+        """A convenience function that initializes a Resampler from a state_dict.
+
+        Some of the shape parameters are inferred from the state_dict (e.g. dim, embedding_dim, etc.). At the time of
+        writing, we did not have a need for inferring ALL of the shape parameters from the state_dict, but this would be
+        possible if needed in the future.
+
+        Args:
+            state_dict (dict[torch.Tensor]): The state_dict to load.
+            depth (int, optional):
+            dim_head (int, optional):
+            heads (int, optional):
+            ff_mult (int, optional):
+
+        Returns:
+            Resampler
+        """
+        dim = state_dict["latents"].shape[2]
+        num_queries = state_dict["latents"].shape[1]
+        embedding_dim = state_dict["proj_in.weight"].shape[-1]
+        output_dim = state_dict["norm_out.weight"].shape[0]
+
+        model = cls(
+            dim=dim,
+            depth=depth,
+            dim_head=dim_head,
+            heads=heads,
+            num_queries=num_queries,
+            embedding_dim=embedding_dim,
+            output_dim=output_dim,
+            ff_mult=ff_mult,
+        )
+        model.load_state_dict(state_dict)
+        return model
+
+    def forward(self, x):
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)

invokeai/backend/ip_adapter/unet_patcher.py

@@ -0,0 +1,53 @@
+from contextlib import contextmanager
+
+from diffusers.models import UNet2DConditionModel
+
+from invokeai.backend.ip_adapter.attention_processor import AttnProcessor2_0, IPAttnProcessor2_0
+from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
+
+
+class UNetPatcher:
+    """A class that contains multiple IP-Adapters and can apply them to a UNet."""
+
+    def __init__(self, ip_adapters: list[IPAdapter]):
+        self._ip_adapters = ip_adapters
+        self._scales = [1.0] * len(self._ip_adapters)
+
+    def set_scale(self, idx: int, value: float):
+        self._scales[idx] = value
+
+    def _prepare_attention_processors(self, unet: UNet2DConditionModel):
+        """Prepare a dict of attention processors that can be injected into a unet, and load the IP-Adapter attention
+        weights into them.
+
+        Note that the `unet` param is only used to determine attention block dimensions and naming.
+        """
+        # Construct a dict of attention processors based on the UNet's architecture.
+        attn_procs = {}
+        for idx, name in enumerate(unet.attn_processors.keys()):
+            if name.endswith("attn1.processor"):
+                attn_procs[name] = AttnProcessor2_0()
+            else:
+                # Collect the weights from each IP Adapter for the idx'th attention processor.
+                attn_procs[name] = IPAttnProcessor2_0(
+                    [ip_adapter.attn_weights.get_attention_processor_weights(idx) for ip_adapter in self._ip_adapters],
+                    self._scales,
+                )
+        return attn_procs
+
+    @contextmanager
+    def apply_ip_adapter_attention(self, unet: UNet2DConditionModel):
+        """A context manager that patches `unet` with IP-Adapter attention processors."""
+
+        attn_procs = self._prepare_attention_processors(unet)
+
+        orig_attn_processors = unet.attn_processors
+
+        try:
+            # Note to future devs: set_attn_processor(...) does something slightly unexpected - it pops elements from the
+            # passed dict. So, if you wanted to keep the dict for future use, you'd have to make a moderately-shallow copy
+            # of it. E.g. `attn_procs_copy = {k: v for k, v in attn_procs.items()}`.
+            unet.set_attn_processor(attn_procs)
+            yield None
+        finally:
+            unet.set_attn_processor(orig_attn_processors)

invokeai/backend/model_management/README.md

@@ -0,0 +1,27 @@
+# Model Cache
+
+## `glibc` Memory Allocator Fragmentation
+
+Python (and PyTorch) relies on the memory allocator from the C Standard Library (`libc`). On linux, with the GNU C Standard Library implementation (`glibc`), our memory access patterns have been observed to cause severe memory fragmentation. This fragmentation results in large amounts of memory that has been freed but can't be released back to the OS. Loading models from disk and moving them between CPU/CUDA seem to be the operations that contribute most to the fragmentation. This memory fragmentation issue can result in OOM crashes during frequent model switching, even if `max_cache_size` is set to a reasonable value (e.g. a OOM crash with `max_cache_size=16` on a system with 32GB of RAM).
+
+This problem may also exist on other OSes, and other `libc` implementations. But, at the time of writing, it has only been investigated on linux with `glibc`.
+
+To better understand how the `glibc` memory allocator works, see these references:
+- Basics: https://www.gnu.org/software/libc/manual/html_node/The-GNU-Allocator.html
+- Details: https://sourceware.org/glibc/wiki/MallocInternals
+
+Note the differences between memory allocated as chunks in an arena vs. memory allocated with `mmap`. Under `glibc`'s default configuration, most model tensors get allocated as chunks in an arena making them vulnerable to the problem of fragmentation.
+
+We can work around this memory fragmentation issue by setting the following env var:
+
+```bash
+# Force blocks >1MB to be allocated with `mmap` so that they are released to the system immediately when they are freed.
+MALLOC_MMAP_THRESHOLD_=1048576
+```
+
+See the following references for more information about the `malloc` tunable parameters:
+- https://www.gnu.org/software/libc/manual/html_node/Malloc-Tunable-Parameters.html
+- https://www.gnu.org/software/libc/manual/html_node/Memory-Allocation-Tunables.html
+- https://man7.org/linux/man-pages/man3/mallopt.3.html
+
+The model cache emits debug logs that provide visibility into the state of the `libc` memory allocator. See the `LibcUtil` class for more info on how these `libc` malloc stats are collected.

invokeai/backend/model_management/convert_ckpt_to_diffusers.py

@@ -74,7 +74,7 @@ if is_accelerate_available():
     from accelerate import init_empty_weights
     from accelerate.utils import set_module_tensor_to_device
 
-logger = InvokeAILogger.getLogger(__name__)
+logger = InvokeAILogger.get_logger(__name__)
 CONVERT_MODEL_ROOT = InvokeAIAppConfig.get_config().models_path / "core/convert"
 
 
@@ -1279,12 +1279,12 @@ def download_from_original_stable_diffusion_ckpt(
         extract_ema = original_config["model"]["params"]["use_ema"]
 
     if (
-        model_version == BaseModelType.StableDiffusion2
+        model_version in [BaseModelType.StableDiffusion2, BaseModelType.StableDiffusion1]
         and original_config["model"]["params"].get("parameterization") == "v"
     ):
         prediction_type = "v_prediction"
         upcast_attention = True
-        image_size = 768
+        image_size = 768 if model_version == BaseModelType.StableDiffusion2 else 512
     else:
         prediction_type = "epsilon"
         upcast_attention = False

invokeai/backend/model_management/libc_util.py

@@ -0,0 +1,75 @@
+import ctypes
+
+
+class Struct_mallinfo2(ctypes.Structure):
+    """A ctypes Structure that matches the libc mallinfo2 struct.
+
+    Docs:
+    - https://man7.org/linux/man-pages/man3/mallinfo.3.html
+    - https://www.gnu.org/software/libc/manual/html_node/Statistics-of-Malloc.html
+
+    struct mallinfo2 {
+        size_t arena;     /* Non-mmapped space allocated (bytes) */
+        size_t ordblks;   /* Number of free chunks */
+        size_t smblks;    /* Number of free fastbin blocks */
+        size_t hblks;     /* Number of mmapped regions */
+        size_t hblkhd;    /* Space allocated in mmapped regions (bytes) */
+        size_t usmblks;   /* See below */
+        size_t fsmblks;   /* Space in freed fastbin blocks (bytes) */
+        size_t uordblks;  /* Total allocated space (bytes) */
+        size_t fordblks;  /* Total free space (bytes) */
+        size_t keepcost;  /* Top-most, releasable space (bytes) */
+    };
+    """
+
+    _fields_ = [
+        ("arena", ctypes.c_size_t),
+        ("ordblks", ctypes.c_size_t),
+        ("smblks", ctypes.c_size_t),
+        ("hblks", ctypes.c_size_t),
+        ("hblkhd", ctypes.c_size_t),
+        ("usmblks", ctypes.c_size_t),
+        ("fsmblks", ctypes.c_size_t),
+        ("uordblks", ctypes.c_size_t),
+        ("fordblks", ctypes.c_size_t),
+        ("keepcost", ctypes.c_size_t),
+    ]
+
+    def __str__(self):
+        s = ""
+        s += f"{'arena': <10}= {(self.arena/2**30):15.5f}   # Non-mmapped space allocated (GB) (uordblks + fordblks)\n"
+        s += f"{'ordblks': <10}= {(self.ordblks): >15}   # Number of free chunks\n"
+        s += f"{'smblks': <10}= {(self.smblks): >15}   # Number of free fastbin blocks \n"
+        s += f"{'hblks': <10}= {(self.hblks): >15}   # Number of mmapped regions \n"
+        s += f"{'hblkhd': <10}= {(self.hblkhd/2**30):15.5f}   # Space allocated in mmapped regions (GB)\n"
+        s += f"{'usmblks': <10}= {(self.usmblks): >15}   # Unused\n"
+        s += f"{'fsmblks': <10}= {(self.fsmblks/2**30):15.5f}   # Space in freed fastbin blocks (GB)\n"
+        s += (
+            f"{'uordblks': <10}= {(self.uordblks/2**30):15.5f}   # Space used by in-use allocations (non-mmapped)"
+            " (GB)\n"
+        )
+        s += f"{'fordblks': <10}= {(self.fordblks/2**30):15.5f}   # Space in free blocks (non-mmapped) (GB)\n"
+        s += f"{'keepcost': <10}= {(self.keepcost/2**30):15.5f}   # Top-most, releasable space (GB)\n"
+        return s
+
+
+class LibcUtil:
+    """A utility class for interacting with the C Standard Library (`libc`) via ctypes.
+
+    Note that this class will raise on __init__() if 'libc.so.6' can't be found. Take care to handle environments where
+    this shared library is not available.
+
+    TODO: Improve cross-OS compatibility of this class.
+    """
+
+    def __init__(self):
+        self._libc = ctypes.cdll.LoadLibrary("libc.so.6")
+
+    def mallinfo2(self) -> Struct_mallinfo2:
+        """Calls `libc` `mallinfo2`.
+
+        Docs: https://man7.org/linux/man-pages/man3/mallinfo.3.html
+        """
+        mallinfo2 = self._libc.mallinfo2
+        mallinfo2.restype = Struct_mallinfo2
+        return mallinfo2()

invokeai/backend/model_management/memory_snapshot.py

@@ -0,0 +1,94 @@
+import gc
+from typing import Optional
+
+import psutil
+import torch
+
+from invokeai.backend.model_management.libc_util import LibcUtil, Struct_mallinfo2
+
+GB = 2**30  # 1 GB
+
+
+class MemorySnapshot:
+    """A snapshot of RAM and VRAM usage. All values are in bytes."""
+
+    def __init__(self, process_ram: int, vram: Optional[int], malloc_info: Optional[Struct_mallinfo2]):
+        """Initialize a MemorySnapshot.
+
+        Most of the time, `MemorySnapshot` will be constructed with `MemorySnapshot.capture()`.
+
+        Args:
+            process_ram (int): CPU RAM used by the current process.
+            vram (Optional[int]): VRAM used by torch.
+            malloc_info (Optional[Struct_mallinfo2]): Malloc info obtained from LibcUtil.
+        """
+        self.process_ram = process_ram
+        self.vram = vram
+        self.malloc_info = malloc_info
+
+    @classmethod
+    def capture(cls, run_garbage_collector: bool = True):
+        """Capture and return a MemorySnapshot.
+
+        Note: This function has significant overhead, particularly if `run_garbage_collector == True`.
+
+        Args:
+            run_garbage_collector (bool, optional): If true, gc.collect() will be run before checking the process RAM
+                usage. Defaults to True.
+
+        Returns:
+            MemorySnapshot
+        """
+        if run_garbage_collector:
+            gc.collect()
+
+        # According to the psutil docs (https://psutil.readthedocs.io/en/latest/#psutil.Process.memory_info), rss is
+        # supported on all platforms.
+        process_ram = psutil.Process().memory_info().rss
+
+        if torch.cuda.is_available():
+            vram = torch.cuda.memory_allocated()
+        else:
+            # TODO: We could add support for mps.current_allocated_memory() as well. Leaving out for now until we have
+            # time to test it properly.
+            vram = None
+
+        try:
+            malloc_info = LibcUtil().mallinfo2()
+        except OSError:
+            # This is expected in environments that do not have the 'libc.so.6' shared library.
+            malloc_info = None
+
+        return cls(process_ram, vram, malloc_info)
+
+
+def get_pretty_snapshot_diff(snapshot_1: MemorySnapshot, snapshot_2: MemorySnapshot) -> str:
+    """Get a pretty string describing the difference between two `MemorySnapshot`s."""
+
+    def get_msg_line(prefix: str, val1: int, val2: int):
+        diff = val2 - val1
+        return f"{prefix: <30} ({(diff/GB):+5.3f}): {(val1/GB):5.3f}GB -> {(val2/GB):5.3f}GB\n"
+
+    msg = ""
+
+    msg += get_msg_line("Process RAM", snapshot_1.process_ram, snapshot_2.process_ram)
+
+    if snapshot_1.malloc_info is not None and snapshot_2.malloc_info is not None:
+        msg += get_msg_line("libc mmap allocated", snapshot_1.malloc_info.hblkhd, snapshot_2.malloc_info.hblkhd)
+
+        msg += get_msg_line("libc arena used", snapshot_1.malloc_info.uordblks, snapshot_2.malloc_info.uordblks)
+
+        msg += get_msg_line("libc arena free", snapshot_1.malloc_info.fordblks, snapshot_2.malloc_info.fordblks)
+
+        libc_total_allocated_1 = snapshot_1.malloc_info.arena + snapshot_1.malloc_info.hblkhd
+        libc_total_allocated_2 = snapshot_2.malloc_info.arena + snapshot_2.malloc_info.hblkhd
+        msg += get_msg_line("libc total allocated", libc_total_allocated_1, libc_total_allocated_2)
+
+        libc_total_used_1 = snapshot_1.malloc_info.uordblks + snapshot_1.malloc_info.hblkhd
+        libc_total_used_2 = snapshot_2.malloc_info.uordblks + snapshot_2.malloc_info.hblkhd
+        msg += get_msg_line("libc total used", libc_total_used_1, libc_total_used_2)
+
+    if snapshot_1.vram is not None and snapshot_2.vram is not None:
+        msg += get_msg_line("VRAM", snapshot_1.vram, snapshot_2.vram)
+
+    return msg

invokeai/backend/model_management/model_cache.py

@@ -18,8 +18,10 @@ context. Use like this:
 
 import gc
 import hashlib
+import math
 import os
 import sys
+import time
 from contextlib import suppress
 from dataclasses import dataclass, field
 from pathlib import Path
@@ -28,6 +30,8 @@ from typing import Any, Dict, Optional, Type, Union, types
 import torch
 
 import invokeai.backend.util.logging as logger
+from invokeai.backend.model_management.memory_snapshot import MemorySnapshot, get_pretty_snapshot_diff
+from invokeai.backend.model_management.model_load_optimizations import skip_torch_weight_init
 
 from ..util.devices import choose_torch_device
 from .models import BaseModelType, ModelBase, ModelType, SubModelType
@@ -44,6 +48,8 @@ DEFAULT_MAX_VRAM_CACHE_SIZE = 2.75
 
 # actual size of a gig
 GIG = 1073741824
+# Size of a MB in bytes.
+MB = 2**20
 
 
 @dataclass
@@ -205,22 +211,41 @@ class ModelCache(object):
         cache_entry = self._cached_models.get(key, None)
         if cache_entry is None:
             self.logger.info(
-                f"Loading model {model_path}, type {base_model.value}:{model_type.value}{':'+submodel.value if submodel else ''}"
+                f"Loading model {model_path}, type"
+                f" {base_model.value}:{model_type.value}{':'+submodel.value if submodel else ''}"
             )
             if self.stats:
                 self.stats.misses += 1
 
-            # this will remove older cached models until
-            # there is sufficient room to load the requested model
-            self._make_cache_room(model_info.get_size(submodel))
+            self_reported_model_size_before_load = model_info.get_size(submodel)
+            # Remove old models from the cache to make room for the new model.
+            self._make_cache_room(self_reported_model_size_before_load)
 
-            # clean memory to make MemoryUsage() more accurate
-            gc.collect()
-            model = model_info.get_model(child_type=submodel, torch_dtype=self.precision)
-            if mem_used := model_info.get_size(submodel):
-                self.logger.debug(f"CPU RAM used for load: {(mem_used/GIG):.2f} GB")
+            # Load the model from disk and capture a memory snapshot before/after.
+            start_load_time = time.time()
+            snapshot_before = MemorySnapshot.capture()
+            with skip_torch_weight_init():
+                model = model_info.get_model(child_type=submodel, torch_dtype=self.precision)
+            snapshot_after = MemorySnapshot.capture()
+            end_load_time = time.time()
 
-            cache_entry = _CacheRecord(self, model, mem_used)
+            self_reported_model_size_after_load = model_info.get_size(submodel)
+
+            self.logger.debug(
+                f"Moved model '{key}' from disk to cpu in {(end_load_time-start_load_time):.2f}s.\n"
+                f"Self-reported size before/after load: {(self_reported_model_size_before_load/GIG):.3f}GB /"
+                f" {(self_reported_model_size_after_load/GIG):.3f}GB.\n"
+                f"{get_pretty_snapshot_diff(snapshot_before, snapshot_after)}"
+            )
+
+            if abs(self_reported_model_size_after_load - self_reported_model_size_before_load) > 10 * MB:
+                self.logger.debug(
+                    f"Model '{key}' mis-reported its size before load. Self-reported size before/after load:"
+                    f" {(self_reported_model_size_before_load/GIG):.2f}GB /"
+                    f" {(self_reported_model_size_after_load/GIG):.2f}GB."
+                )
+
+            cache_entry = _CacheRecord(self, model, self_reported_model_size_after_load)
             self._cached_models[key] = cache_entry
         else:
             if self.stats:
@@ -240,6 +265,45 @@ class ModelCache(object):
 
         return self.ModelLocker(self, key, cache_entry.model, gpu_load, cache_entry.size)
 
+    def _move_model_to_device(self, key: str, target_device: torch.device):
+        cache_entry = self._cached_models[key]
+
+        source_device = cache_entry.model.device
+        # Note: We compare device types only so that 'cuda' == 'cuda:0'. This would need to be revised to support
+        # multi-GPU.
+        if torch.device(source_device).type == torch.device(target_device).type:
+            return
+
+        start_model_to_time = time.time()
+        snapshot_before = MemorySnapshot.capture()
+        cache_entry.model.to(target_device)
+        snapshot_after = MemorySnapshot.capture()
+        end_model_to_time = time.time()
+        self.logger.debug(
+            f"Moved model '{key}' from {source_device} to"
+            f" {target_device} in {(end_model_to_time-start_model_to_time):.2f}s.\n"
+            f"Estimated model size: {(cache_entry.size/GIG):.3f} GB.\n"
+            f"{get_pretty_snapshot_diff(snapshot_before, snapshot_after)}"
+        )
+
+        if snapshot_before.vram is not None and snapshot_after.vram is not None:
+            vram_change = abs(snapshot_before.vram - snapshot_after.vram)
+
+            # If the estimated model size does not match the change in VRAM, log a warning.
+            if not math.isclose(
+                vram_change,
+                cache_entry.size,
+                rel_tol=0.1,
+                abs_tol=10 * MB,
+            ):
+                self.logger.debug(
+                    f"Moving model '{key}' from {source_device} to"
+                    f" {target_device} caused an unexpected change in VRAM usage. The model's"
+                    " estimated size may be incorrect. Estimated model size:"
+                    f" {(cache_entry.size/GIG):.3f} GB.\n"
+                    f"{get_pretty_snapshot_diff(snapshot_before, snapshot_after)}"
+                )
+
     class ModelLocker(object):
         def __init__(self, cache, key, model, gpu_load, size_needed):
             """
@@ -269,11 +333,7 @@ class ModelCache(object):
                     if self.cache.lazy_offloading:
                         self.cache._offload_unlocked_models(self.size_needed)
 
-                    if self.model.device != self.cache.execution_device:
-                        self.cache.logger.debug(f"Moving {self.key} into {self.cache.execution_device}")
-                        with VRAMUsage() as mem:
-                            self.model.to(self.cache.execution_device)  # move into GPU
-                        self.cache.logger.debug(f"GPU VRAM used for load: {(mem.vram_used/GIG):.2f} GB")
+                    self.cache._move_model_to_device(self.key, self.cache.execution_device)
 
                     self.cache.logger.debug(f"Locking {self.key} in {self.cache.execution_device}")
                     self.cache._print_cuda_stats()
@@ -286,7 +346,7 @@ class ModelCache(object):
             # in the event that the caller wants the model in RAM, we
             # move it into CPU if it is in GPU and not locked
             elif self.cache_entry.loaded and not self.cache_entry.locked:
-                self.model.to(self.cache.storage_device)
+                self.cache._move_model_to_device(self.key, self.cache.storage_device)
 
             return self.model
 
@@ -339,7 +399,8 @@ class ModelCache(object):
                 locked_models += 1
 
         self.logger.debug(
-            f"Current VRAM/RAM usage: {vram}/{ram}; cached_models/loaded_models/locked_models/ = {cached_models}/{loaded_models}/{locked_models}"
+            f"Current VRAM/RAM usage: {vram}/{ram}; cached_models/loaded_models/locked_models/ ="
+            f" {cached_models}/{loaded_models}/{locked_models}"
         )
 
     def _cache_size(self) -> int:
@@ -354,7 +415,8 @@ class ModelCache(object):
 
         if current_size + bytes_needed > maximum_size:
             self.logger.debug(
-                f"Max cache size exceeded: {(current_size/GIG):.2f}/{self.max_cache_size:.2f} GB, need an additional {(bytes_needed/GIG):.2f} GB"
+                f"Max cache size exceeded: {(current_size/GIG):.2f}/{self.max_cache_size:.2f} GB, need an additional"
+                f" {(bytes_needed/GIG):.2f} GB"
             )
 
         self.logger.debug(f"Before unloading: cached_models={len(self._cached_models)}")
@@ -387,7 +449,8 @@ class ModelCache(object):
 
             device = cache_entry.model.device if hasattr(cache_entry.model, "device") else None
             self.logger.debug(
-                f"Model: {model_key}, locks: {cache_entry._locks}, device: {device}, loaded: {cache_entry.loaded}, refs: {refs}"
+                f"Model: {model_key}, locks: {cache_entry._locks}, device: {device}, loaded: {cache_entry.loaded},"
+                f" refs: {refs}"
             )
 
             # 2 refs:
@@ -423,11 +486,9 @@ class ModelCache(object):
             if vram_in_use <= reserved:
                 break
             if not cache_entry.locked and cache_entry.loaded:
-                self.logger.debug(f"Offloading {model_key} from {self.execution_device} into {self.storage_device}")
-                with VRAMUsage() as mem:
-                    cache_entry.model.to(self.storage_device)
-                self.logger.debug(f"GPU VRAM freed: {(mem.vram_used/GIG):.2f} GB")
-                vram_in_use += mem.vram_used  # note vram_used is negative
+                self._move_model_to_device(model_key, self.storage_device)
+
+                vram_in_use = torch.cuda.memory_allocated()
                 self.logger.debug(f"{(vram_in_use/GIG):.2f}GB VRAM used for models; max allowed={(reserved/GIG):.2f}GB")
 
         gc.collect()
@@ -454,16 +515,3 @@ class ModelCache(object):
         with open(hashpath, "w") as f:
             f.write(hash)
         return hash
-
-
-class VRAMUsage(object):
-    def __init__(self):
-        self.vram = None
-        self.vram_used = 0
-
-    def __enter__(self):
-        self.vram = torch.cuda.memory_allocated()
-        return self
-
-    def __exit__(self, *args):
-        self.vram_used = torch.cuda.memory_allocated() - self.vram

invokeai/backend/model_management/model_load_optimizations.py

@@ -0,0 +1,30 @@
+from contextlib import contextmanager
+
+import torch
+
+
+def _no_op(*args, **kwargs):
+    pass
+
+
+@contextmanager
+def skip_torch_weight_init():
+    """A context manager that monkey-patches several of the common torch layers (torch.nn.Linear, torch.nn.Conv1d, etc.)
+    to skip weight initialization.
+
+    By default, `torch.nn.Linear` and `torch.nn.ConvNd` layers initialize their weights (according to a particular
+    distribution) when __init__ is called. This weight initialization step can take a significant amount of time, and is
+    completely unnecessary if the intent is to load checkpoint weights from disk for the layer. This context manager
+    monkey-patches common torch layers to skip the weight initialization step.
+    """
+    torch_modules = [torch.nn.Linear, torch.nn.modules.conv._ConvNd]
+    saved_functions = [m.reset_parameters for m in torch_modules]
+
+    try:
+        for torch_module in torch_modules:
+            torch_module.reset_parameters = _no_op
+
+        yield None
+    finally:
+        for torch_module, saved_function in zip(torch_modules, saved_functions):
+            torch_module.reset_parameters = saved_function

invokeai/backend/model_management/model_manager.py

@@ -25,6 +25,7 @@ Models are described using four attributes:
          ModelType.Lora -- a LoRA or LyCORIS fine-tune
          ModelType.TextualInversion -- a textual inversion embedding
          ModelType.ControlNet -- a ControlNet model
+         ModelType.IPAdapter -- an IPAdapter model
 
   3) BaseModelType -- an enum indicating the stable diffusion base model, one of:
          BaseModelType.StableDiffusion1
@@ -1000,8 +1001,8 @@ class ModelManager(object):
                                 new_models_found = True
                             except DuplicateModelException as e:
                                 self.logger.warning(e)
-                            except InvalidModelException:
-                                self.logger.warning(f"Not a valid model: {model_path}")
+                            except InvalidModelException as e:
+                                self.logger.warning(f"Not a valid model: {model_path}. {e}")
                             except NotImplementedError as e:
                                 self.logger.warning(e)
 

invokeai/backend/model_management/model_probe.py

@@ -1,4 +1,5 @@
 import json
+import re
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Callable, Dict, Literal, Optional, Union
@@ -8,6 +9,8 @@ import torch
 from diffusers import ConfigMixin, ModelMixin
 from picklescan.scanner import scan_file_path
 
+from invokeai.backend.model_management.models.ip_adapter import IPAdapterModelFormat
+
 from .models import (
     BaseModelType,
     InvalidModelException,
@@ -51,7 +54,10 @@ class ModelProbe(object):
         "StableDiffusionXLImg2ImgPipeline": ModelType.Main,
         "StableDiffusionXLInpaintPipeline": ModelType.Main,
         "AutoencoderKL": ModelType.Vae,
+        "AutoencoderTiny": ModelType.Vae,
         "ControlNetModel": ModelType.ControlNet,
+        "CLIPVisionModelWithProjection": ModelType.CLIPVision,
+        "T2IAdapter": ModelType.T2IAdapter,
     }
 
     @classmethod
@@ -85,8 +91,7 @@ class ModelProbe(object):
         to place it somewhere in the models directory hierarchy. If the model is
         already loaded into memory, you may provide it as model in order to avoid
         opening it a second time. The prediction_type_helper callable is a function that receives
-        the path to the model and returns the BaseModelType. It is called to distinguish
-        between V2-Base and V2-768 SD models.
+        the path to the model and returns the SchedulerPredictionType.
         """
         if model_path:
             format_type = "diffusers" if model_path.is_dir() else "checkpoint"
@@ -118,14 +123,18 @@ class ModelProbe(object):
                     and prediction_type == SchedulerPredictionType.VPrediction
                 ),
                 format=format,
-                image_size=1024
-                if (base_type in {BaseModelType.StableDiffusionXL, BaseModelType.StableDiffusionXLRefiner})
-                else 768
-                if (
-                    base_type == BaseModelType.StableDiffusion2
-                    and prediction_type == SchedulerPredictionType.VPrediction
-                )
-                else 512,
+                image_size=(
+                    1024
+                    if (base_type in {BaseModelType.StableDiffusionXL, BaseModelType.StableDiffusionXLRefiner})
+                    else (
+                        768
+                        if (
+                            base_type == BaseModelType.StableDiffusion2
+                            and prediction_type == SchedulerPredictionType.VPrediction
+                        )
+                        else 512
+                    )
+                ),
             )
         except Exception:
             raise
@@ -170,6 +179,7 @@ class ModelProbe(object):
         Get the model type of a hugging-face style folder.
         """
         class_name = None
+        error_hint = None
         if model:
             class_name = model.__class__.__name__
         else:
@@ -177,9 +187,10 @@ class ModelProbe(object):
                 return ModelType.ONNX
             if (folder_path / "learned_embeds.bin").exists():
                 return ModelType.TextualInversion
-
             if (folder_path / "pytorch_lora_weights.bin").exists():
                 return ModelType.Lora
+            if (folder_path / "image_encoder.txt").exists():
+                return ModelType.IPAdapter
 
             i = folder_path / "model_index.json"
             c = folder_path / "config.json"
@@ -188,13 +199,24 @@ class ModelProbe(object):
             if config_path:
                 with open(config_path, "r") as file:
                     conf = json.load(file)
-                class_name = conf["_class_name"]
+                if "_class_name" in conf:
+                    class_name = conf["_class_name"]
+                elif "architectures" in conf:
+                    class_name = conf["architectures"][0]
+                else:
+                    class_name = None
+            else:
+                error_hint = f"No model_index.json or config.json found in {folder_path}."
 
         if class_name and (type := cls.CLASS2TYPE.get(class_name)):
             return type
+        else:
+            error_hint = f"class {class_name} is not one of the supported classes [{', '.join(cls.CLASS2TYPE.keys())}]"
 
         # give up
-        raise InvalidModelException(f"Unable to determine model type for {folder_path}")
+        raise InvalidModelException(
+            f"Unable to determine model type for {folder_path}" + (f"; {error_hint}" if error_hint else "")
+        )
 
     @classmethod
     def _scan_and_load_checkpoint(cls, model_path: Path) -> dict:
@@ -283,25 +305,36 @@ class PipelineCheckpointProbe(CheckpointProbeBase):
         else:
             raise InvalidModelException("Cannot determine base type")
 
-    def get_scheduler_prediction_type(self) -> SchedulerPredictionType:
+    def get_scheduler_prediction_type(self) -> Optional[SchedulerPredictionType]:
+        """Return model prediction type."""
+        # if there is a .yaml associated with this checkpoint, then we do not need
+        # to probe for the prediction type as it will be ignored.
+        if self.checkpoint_path and self.checkpoint_path.with_suffix(".yaml").exists():
+            return None
+
         type = self.get_base_type()
-        if type == BaseModelType.StableDiffusion1:
-            return SchedulerPredictionType.Epsilon
-        checkpoint = self.checkpoint
-        state_dict = self.checkpoint.get("state_dict") or checkpoint
-        key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
-        if key_name in state_dict and state_dict[key_name].shape[-1] == 1024:
-            if "global_step" in checkpoint:
-                if checkpoint["global_step"] == 220000:
-                    return SchedulerPredictionType.Epsilon
-                elif checkpoint["global_step"] == 110000:
-                    return SchedulerPredictionType.VPrediction
-            if (
-                self.checkpoint_path and self.helper and not self.checkpoint_path.with_suffix(".yaml").exists()
-            ):  # if a .yaml config file exists, then this step not needed
-                return self.helper(self.checkpoint_path)
-            else:
-                return None
+        if type == BaseModelType.StableDiffusion2:
+            checkpoint = self.checkpoint
+            state_dict = self.checkpoint.get("state_dict") or checkpoint
+            key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
+            if key_name in state_dict and state_dict[key_name].shape[-1] == 1024:
+                if "global_step" in checkpoint:
+                    if checkpoint["global_step"] == 220000:
+                        return SchedulerPredictionType.Epsilon
+                    elif checkpoint["global_step"] == 110000:
+                        return SchedulerPredictionType.VPrediction
+            if self.helper and self.checkpoint_path:
+                if helper_guess := self.helper(self.checkpoint_path):
+                    return helper_guess
+            return SchedulerPredictionType.VPrediction  # a guess for sd2 ckpts
+
+        elif type == BaseModelType.StableDiffusion1:
+            if self.helper and self.checkpoint_path:
+                if helper_guess := self.helper(self.checkpoint_path):
+                    return helper_guess
+            return SchedulerPredictionType.Epsilon  # a reasonable guess for sd1 ckpts
+        else:
+            return None
 
 
 class VaeCheckpointProbe(CheckpointProbeBase):
@@ -366,6 +399,21 @@ class ControlNetCheckpointProbe(CheckpointProbeBase):
         raise InvalidModelException("Unable to determine base type for {self.checkpoint_path}")
 
 
+class IPAdapterCheckpointProbe(CheckpointProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        raise NotImplementedError()
+
+
+class CLIPVisionCheckpointProbe(CheckpointProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        raise NotImplementedError()
+
+
+class T2IAdapterCheckpointProbe(CheckpointProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        raise NotImplementedError()
+
+
 ########################################################
 # classes for probing folders
 #######################################################
@@ -438,16 +486,32 @@ class PipelineFolderProbe(FolderProbeBase):
 
 class VaeFolderProbe(FolderProbeBase):
     def get_base_type(self) -> BaseModelType:
+        if self._config_looks_like_sdxl():
+            return BaseModelType.StableDiffusionXL
+        elif self._name_looks_like_sdxl():
+            # but SD and SDXL VAE are the same shape (3-channel RGB to 4-channel float scaled down
+            # by a factor of 8), we can't necessarily tell them apart by config hyperparameters.
+            return BaseModelType.StableDiffusionXL
+        else:
+            return BaseModelType.StableDiffusion1
+
+    def _config_looks_like_sdxl(self) -> bool:
+        # config values that distinguish Stability's SD 1.x VAE from their SDXL VAE.
         config_file = self.folder_path / "config.json"
         if not config_file.exists():
             raise InvalidModelException(f"Cannot determine base type for {self.folder_path}")
         with open(config_file, "r") as file:
             config = json.load(file)
-        return (
-            BaseModelType.StableDiffusionXL
-            if config.get("scaling_factor", 0) == 0.13025 and config.get("sample_size") in [512, 1024]
-            else BaseModelType.StableDiffusion1
-        )
+        return config.get("scaling_factor", 0) == 0.13025 and config.get("sample_size") in [512, 1024]
+
+    def _name_looks_like_sdxl(self) -> bool:
+        return bool(re.search(r"xl\b", self._guess_name(), re.IGNORECASE))
+
+    def _guess_name(self) -> str:
+        name = self.folder_path.name
+        if name == "vae":
+            name = self.folder_path.parent.name
+        return name
 
 
 class TextualInversionFolderProbe(FolderProbeBase):
@@ -485,11 +549,13 @@ class ControlNetFolderProbe(FolderProbeBase):
         base_model = (
             BaseModelType.StableDiffusion1
             if dimension == 768
-            else BaseModelType.StableDiffusion2
-            if dimension == 1024
-            else BaseModelType.StableDiffusionXL
-            if dimension == 2048
-            else None
+            else (
+                BaseModelType.StableDiffusion2
+                if dimension == 1024
+                else BaseModelType.StableDiffusionXL
+                if dimension == 2048
+                else None
+            )
         )
         if not base_model:
             raise InvalidModelException(f"Unable to determine model base for {self.folder_path}")
@@ -509,15 +575,69 @@ class LoRAFolderProbe(FolderProbeBase):
         return LoRACheckpointProbe(model_file, None).get_base_type()
 
 
+class IPAdapterFolderProbe(FolderProbeBase):
+    def get_format(self) -> str:
+        return IPAdapterModelFormat.InvokeAI.value
+
+    def get_base_type(self) -> BaseModelType:
+        model_file = self.folder_path / "ip_adapter.bin"
+        if not model_file.exists():
+            raise InvalidModelException("Unknown IP-Adapter model format.")
+
+        state_dict = torch.load(model_file, map_location="cpu")
+        cross_attention_dim = state_dict["ip_adapter"]["1.to_k_ip.weight"].shape[-1]
+        if cross_attention_dim == 768:
+            return BaseModelType.StableDiffusion1
+        elif cross_attention_dim == 1024:
+            return BaseModelType.StableDiffusion2
+        elif cross_attention_dim == 2048:
+            return BaseModelType.StableDiffusionXL
+        else:
+            raise InvalidModelException(f"IP-Adapter had unexpected cross-attention dimension: {cross_attention_dim}.")
+
+
+class CLIPVisionFolderProbe(FolderProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        return BaseModelType.Any
+
+
+class T2IAdapterFolderProbe(FolderProbeBase):
+    def get_base_type(self) -> BaseModelType:
+        config_file = self.folder_path / "config.json"
+        if not config_file.exists():
+            raise InvalidModelException(f"Cannot determine base type for {self.folder_path}")
+        with open(config_file, "r") as file:
+            config = json.load(file)
+
+        adapter_type = config.get("adapter_type", None)
+        if adapter_type == "full_adapter_xl":
+            return BaseModelType.StableDiffusionXL
+        elif adapter_type == "full_adapter" or "light_adapter":
+            # I haven't seen any T2I adapter models for SD2, so assume that this is an SD1 adapter.
+            return BaseModelType.StableDiffusion1
+        else:
+            raise InvalidModelException(
+                f"Unable to determine base model for '{self.folder_path}' (adapter_type = {adapter_type})."
+            )
+
+
 ############## register probe classes ######
 ModelProbe.register_probe("diffusers", ModelType.Main, PipelineFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.Vae, VaeFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.Lora, LoRAFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.TextualInversion, TextualInversionFolderProbe)
 ModelProbe.register_probe("diffusers", ModelType.ControlNet, ControlNetFolderProbe)
+ModelProbe.register_probe("diffusers", ModelType.IPAdapter, IPAdapterFolderProbe)
+ModelProbe.register_probe("diffusers", ModelType.CLIPVision, CLIPVisionFolderProbe)
+ModelProbe.register_probe("diffusers", ModelType.T2IAdapter, T2IAdapterFolderProbe)
+
 ModelProbe.register_probe("checkpoint", ModelType.Main, PipelineCheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.Vae, VaeCheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.Lora, LoRACheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.TextualInversion, TextualInversionCheckpointProbe)
 ModelProbe.register_probe("checkpoint", ModelType.ControlNet, ControlNetCheckpointProbe)
+ModelProbe.register_probe("checkpoint", ModelType.IPAdapter, IPAdapterCheckpointProbe)
+ModelProbe.register_probe("checkpoint", ModelType.CLIPVision, CLIPVisionCheckpointProbe)
+ModelProbe.register_probe("checkpoint", ModelType.T2IAdapter, T2IAdapterCheckpointProbe)
+
 ModelProbe.register_probe("onnx", ModelType.ONNX, ONNXFolderProbe)

invokeai/backend/model_management/model_search.py

@@ -71,7 +71,13 @@ class ModelSearch(ABC):
                 if any(
                     [
                         (path / x).exists()
-                        for x in {"config.json", "model_index.json", "learned_embeds.bin", "pytorch_lora_weights.bin"}
+                        for x in {
+                            "config.json",
+                            "model_index.json",
+                            "learned_embeds.bin",
+                            "pytorch_lora_weights.bin",
+                            "image_encoder.txt",
+                        }
                     ]
                 ):
                     try:
@@ -79,7 +85,7 @@ class ModelSearch(ABC):
                         self._models_found += 1
                         self._scanned_dirs.add(path)
                     except Exception as e:
-                        self.logger.warning(str(e))
+                        self.logger.warning(f"Failed to process '{path}': {e}")
 
             for f in files:
                 path = Path(root) / f
@@ -90,7 +96,7 @@ class ModelSearch(ABC):
                         self.on_model_found(path)
                         self._models_found += 1
                     except Exception as e:
-                        self.logger.warning(str(e))
+                        self.logger.warning(f"Failed to process '{path}': {e}")
 
 
 class FindModels(ModelSearch):

invokeai/backend/model_management/models/__init__.py

@@ -18,11 +18,14 @@ from .base import (  # noqa: F401
     SilenceWarnings,
     SubModelType,
 )
+from .clip_vision import CLIPVisionModel
 from .controlnet import ControlNetModel  # TODO:
+from .ip_adapter import IPAdapterModel
 from .lora import LoRAModel
 from .sdxl import StableDiffusionXLModel
 from .stable_diffusion import StableDiffusion1Model, StableDiffusion2Model
 from .stable_diffusion_onnx import ONNXStableDiffusion1Model, ONNXStableDiffusion2Model
+from .t2i_adapter import T2IAdapterModel
 from .textual_inversion import TextualInversionModel
 from .vae import VaeModel
 
@@ -34,6 +37,9 @@ MODEL_CLASSES = {
         ModelType.Lora: LoRAModel,
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
+        ModelType.T2IAdapter: T2IAdapterModel,
     },
     BaseModelType.StableDiffusion2: {
         ModelType.ONNX: ONNXStableDiffusion2Model,
@@ -42,6 +48,9 @@ MODEL_CLASSES = {
         ModelType.Lora: LoRAModel,
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
+        ModelType.T2IAdapter: T2IAdapterModel,
     },
     BaseModelType.StableDiffusionXL: {
         ModelType.Main: StableDiffusionXLModel,
@@ -51,6 +60,9 @@ MODEL_CLASSES = {
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
         ModelType.ONNX: ONNXStableDiffusion2Model,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
+        ModelType.T2IAdapter: T2IAdapterModel,
     },
     BaseModelType.StableDiffusionXLRefiner: {
         ModelType.Main: StableDiffusionXLModel,
@@ -60,6 +72,21 @@ MODEL_CLASSES = {
         ModelType.ControlNet: ControlNetModel,
         ModelType.TextualInversion: TextualInversionModel,
         ModelType.ONNX: ONNXStableDiffusion2Model,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.CLIPVision: CLIPVisionModel,
+        ModelType.T2IAdapter: T2IAdapterModel,
+    },
+    BaseModelType.Any: {
+        ModelType.CLIPVision: CLIPVisionModel,
+        # The following model types are not expected to be used with BaseModelType.Any.
+        ModelType.ONNX: ONNXStableDiffusion2Model,
+        ModelType.Main: StableDiffusion2Model,
+        ModelType.Vae: VaeModel,
+        ModelType.Lora: LoRAModel,
+        ModelType.ControlNet: ControlNetModel,
+        ModelType.TextualInversion: TextualInversionModel,
+        ModelType.IPAdapter: IPAdapterModel,
+        ModelType.T2IAdapter: T2IAdapterModel,
     },
     # BaseModelType.Kandinsky2_1: {
     #    ModelType.Main: Kandinsky2_1Model,

invokeai/backend/model_management/models/base.py

@@ -36,6 +36,7 @@ class ModelNotFoundException(Exception):
 
 
 class BaseModelType(str, Enum):
+    Any = "any"  # For models that are not associated with any particular base model.
     StableDiffusion1 = "sd-1"
     StableDiffusion2 = "sd-2"
     StableDiffusionXL = "sdxl"
@@ -50,6 +51,9 @@ class ModelType(str, Enum):
     Lora = "lora"
     ControlNet = "controlnet"  # used by model_probe
     TextualInversion = "embedding"
+    IPAdapter = "ip_adapter"
+    CLIPVision = "clip_vision"
+    T2IAdapter = "t2i_adapter"
 
 
 class SubModelType(str, Enum):

invokeai/backend/model_management/models/clip_vision.py

@@ -0,0 +1,82 @@
+import os
+from enum import Enum
+from typing import Literal, Optional
+
+import torch
+from transformers import CLIPVisionModelWithProjection
+
+from invokeai.backend.model_management.models.base import (
+    BaseModelType,
+    InvalidModelException,
+    ModelBase,
+    ModelConfigBase,
+    ModelType,
+    SubModelType,
+    calc_model_size_by_data,
+    calc_model_size_by_fs,
+    classproperty,
+)
+
+
+class CLIPVisionModelFormat(str, Enum):
+    Diffusers = "diffusers"
+
+
+class CLIPVisionModel(ModelBase):
+    class DiffusersConfig(ModelConfigBase):
+        model_format: Literal[CLIPVisionModelFormat.Diffusers]
+
+    def __init__(self, model_path: str, base_model: BaseModelType, model_type: ModelType):
+        assert model_type == ModelType.CLIPVision
+        super().__init__(model_path, base_model, model_type)
+
+        self.model_size = calc_model_size_by_fs(self.model_path)
+
+    @classmethod
+    def detect_format(cls, path: str) -> str:
+        if not os.path.exists(path):
+            raise ModuleNotFoundError(f"No CLIP Vision model at path '{path}'.")
+
+        if os.path.isdir(path) and os.path.exists(os.path.join(path, "config.json")):
+            return CLIPVisionModelFormat.Diffusers
+
+        raise InvalidModelException(f"Unexpected CLIP Vision model format: {path}")
+
+    @classproperty
+    def save_to_config(cls) -> bool:
+        return True
+
+    def get_size(self, child_type: Optional[SubModelType] = None) -> int:
+        if child_type is not None:
+            raise ValueError("There are no child models in a CLIP Vision model.")
+
+        return self.model_size
+
+    def get_model(
+        self,
+        torch_dtype: Optional[torch.dtype],
+        child_type: Optional[SubModelType] = None,
+    ) -> CLIPVisionModelWithProjection:
+        if child_type is not None:
+            raise ValueError("There are no child models in a CLIP Vision model.")
+
+        model = CLIPVisionModelWithProjection.from_pretrained(self.model_path, torch_dtype=torch_dtype)
+
+        # Calculate a more accurate model size.
+        self.model_size = calc_model_size_by_data(model)
+
+        return model
+
+    @classmethod
+    def convert_if_required(
+        cls,
+        model_path: str,
+        output_path: str,
+        config: ModelConfigBase,
+        base_model: BaseModelType,
+    ) -> str:
+        format = cls.detect_format(model_path)
+        if format == CLIPVisionModelFormat.Diffusers:
+            return model_path
+        else:
+            raise ValueError(f"Unsupported format: '{format}'.")

invokeai/backend/model_management/models/ip_adapter.py

@@ -0,0 +1,96 @@
+import os
+import typing
+from enum import Enum
+from typing import Literal, Optional
+
+import torch
+
+from invokeai.backend.ip_adapter.ip_adapter import IPAdapter, IPAdapterPlus, build_ip_adapter
+from invokeai.backend.model_management.models.base import (
+    BaseModelType,
+    InvalidModelException,
+    ModelBase,
+    ModelConfigBase,
+    ModelType,
+    SubModelType,
+    calc_model_size_by_fs,
+    classproperty,
+)
+
+
+class IPAdapterModelFormat(str, Enum):
+    # The custom IP-Adapter model format defined by InvokeAI.
+    InvokeAI = "invokeai"
+
+
+class IPAdapterModel(ModelBase):
+    class InvokeAIConfig(ModelConfigBase):
+        model_format: Literal[IPAdapterModelFormat.InvokeAI]
+
+    def __init__(self, model_path: str, base_model: BaseModelType, model_type: ModelType):
+        assert model_type == ModelType.IPAdapter
+        super().__init__(model_path, base_model, model_type)
+
+        self.model_size = calc_model_size_by_fs(self.model_path)
+
+    @classmethod
+    def detect_format(cls, path: str) -> str:
+        if not os.path.exists(path):
+            raise ModuleNotFoundError(f"No IP-Adapter model at path '{path}'.")
+
+        if os.path.isdir(path):
+            model_file = os.path.join(path, "ip_adapter.bin")
+            image_encoder_config_file = os.path.join(path, "image_encoder.txt")
+            if os.path.exists(model_file) and os.path.exists(image_encoder_config_file):
+                return IPAdapterModelFormat.InvokeAI
+
+        raise InvalidModelException(f"Unexpected IP-Adapter model format: {path}")
+
+    @classproperty
+    def save_to_config(cls) -> bool:
+        return True
+
+    def get_size(self, child_type: Optional[SubModelType] = None) -> int:
+        if child_type is not None:
+            raise ValueError("There are no child models in an IP-Adapter model.")
+
+        return self.model_size
+
+    def get_model(
+        self,
+        torch_dtype: Optional[torch.dtype],
+        child_type: Optional[SubModelType] = None,
+    ) -> typing.Union[IPAdapter, IPAdapterPlus]:
+        if child_type is not None:
+            raise ValueError("There are no child models in an IP-Adapter model.")
+
+        model = build_ip_adapter(
+            ip_adapter_ckpt_path=os.path.join(self.model_path, "ip_adapter.bin"), device="cpu", dtype=torch_dtype
+        )
+
+        self.model_size = model.calc_size()
+        return model
+
+    @classmethod
+    def convert_if_required(
+        cls,
+        model_path: str,
+        output_path: str,
+        config: ModelConfigBase,
+        base_model: BaseModelType,
+    ) -> str:
+        format = cls.detect_format(model_path)
+        if format == IPAdapterModelFormat.InvokeAI:
+            return model_path
+        else:
+            raise ValueError(f"Unsupported format: '{format}'.")
+
+
+def get_ip_adapter_image_encoder_model_id(model_path: str):
+    """Read the ID of the image encoder associated with the IP-Adapter at `model_path`."""
+    image_encoder_config_file = os.path.join(model_path, "image_encoder.txt")
+
+    with open(image_encoder_config_file, "r") as f:
+        image_encoder_model = f.readline().strip()
+
+    return image_encoder_model

invokeai/backend/model_management/models/t2i_adapter.py

@@ -0,0 +1,102 @@
+import os
+from enum import Enum
+from typing import Literal, Optional
+
+import torch
+from diffusers import T2IAdapter
+
+from invokeai.backend.model_management.models.base import (
+    BaseModelType,
+    EmptyConfigLoader,
+    InvalidModelException,
+    ModelBase,
+    ModelConfigBase,
+    ModelNotFoundException,
+    ModelType,
+    SubModelType,
+    calc_model_size_by_data,
+    calc_model_size_by_fs,
+    classproperty,
+)
+
+
+class T2IAdapterModelFormat(str, Enum):
+    Diffusers = "diffusers"
+
+
+class T2IAdapterModel(ModelBase):
+    class DiffusersConfig(ModelConfigBase):
+        model_format: Literal[T2IAdapterModelFormat.Diffusers]
+
+    def __init__(self, model_path: str, base_model: BaseModelType, model_type: ModelType):
+        assert model_type == ModelType.T2IAdapter
+        super().__init__(model_path, base_model, model_type)
+
+        config = EmptyConfigLoader.load_config(self.model_path, config_name="config.json")
+
+        model_class_name = config.get("_class_name", None)
+        if model_class_name not in {"T2IAdapter"}:
+            raise InvalidModelException(f"Invalid T2I-Adapter model. Unknown _class_name: '{model_class_name}'.")
+
+        self.model_class = self._hf_definition_to_type(["diffusers", model_class_name])
+        self.model_size = calc_model_size_by_fs(self.model_path)
+
+    def get_size(self, child_type: Optional[SubModelType] = None):
+        if child_type is not None:
+            raise ValueError(f"T2I-Adapters do not have child models. Invalid child type: '{child_type}'.")
+        return self.model_size
+
+    def get_model(
+        self,
+        torch_dtype: Optional[torch.dtype],
+        child_type: Optional[SubModelType] = None,
+    ) -> T2IAdapter:
+        if child_type is not None:
+            raise ValueError(f"T2I-Adapters do not have child models. Invalid child type: '{child_type}'.")
+
+        model = None
+        for variant in ["fp16", None]:
+            try:
+                model = self.model_class.from_pretrained(
+                    self.model_path,
+                    torch_dtype=torch_dtype,
+                    variant=variant,
+                )
+                break
+            except Exception:
+                pass
+        if not model:
+            raise ModelNotFoundException()
+
+        # Calculate a more accurate size after loading the model into memory.
+        self.model_size = calc_model_size_by_data(model)
+        return model
+
+    @classproperty
+    def save_to_config(cls) -> bool:
+        return False
+
+    @classmethod
+    def detect_format(cls, path: str):
+        if not os.path.exists(path):
+            raise ModelNotFoundException(f"Model not found at '{path}'.")
+
+        if os.path.isdir(path):
+            if os.path.exists(os.path.join(path, "config.json")):
+                return T2IAdapterModelFormat.Diffusers
+
+        raise InvalidModelException(f"Unsupported T2I-Adapter format: '{path}'.")
+
+    @classmethod
+    def convert_if_required(
+        cls,
+        model_path: str,
+        output_path: str,
+        config: ModelConfigBase,
+        base_model: BaseModelType,
+    ) -> str:
+        format = cls.detect_format(model_path)
+        if format == T2IAdapterModelFormat.Diffusers:
+            return model_path
+        else:
+            raise ValueError(f"Unsupported format: '{format}'.")

invokeai/backend/stable_diffusion/__init__.py

@@ -1,15 +1,6 @@
 """
 Initialization file for the invokeai.backend.stable_diffusion package
 """
-from .diffusers_pipeline import (  # noqa: F401
-    ConditioningData,
-    PipelineIntermediateState,
-    StableDiffusionGeneratorPipeline,
-)
+from .diffusers_pipeline import PipelineIntermediateState, StableDiffusionGeneratorPipeline  # noqa: F401
 from .diffusion import InvokeAIDiffuserComponent  # noqa: F401
 from .diffusion.cross_attention_map_saving import AttentionMapSaver  # noqa: F401
-from .diffusion.shared_invokeai_diffusion import (  # noqa: F401
-    BasicConditioningInfo,
-    PostprocessingSettings,
-    SDXLConditioningInfo,
-)

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -1,8 +1,8 @@
 from __future__ import annotations
 
-import dataclasses
-import inspect
-from dataclasses import dataclass, field
+import math
+from contextlib import nullcontext
+from dataclasses import dataclass
 from typing import Any, Callable, List, Optional, Union
 
 import einops
@@ -23,9 +23,12 @@ from pydantic import Field
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 
 from invokeai.app.services.config import InvokeAIAppConfig
+from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
+from invokeai.backend.ip_adapter.unet_patcher import UNetPatcher
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import ConditioningData
 
 from ..util import auto_detect_slice_size, normalize_device
-from .diffusion import AttentionMapSaver, BasicConditioningInfo, InvokeAIDiffuserComponent, PostprocessingSettings
+from .diffusion import AttentionMapSaver, InvokeAIDiffuserComponent
 
 
 @dataclass
@@ -95,7 +98,7 @@ class AddsMaskGuidance:
         # Mask anything that has the same shape as prev_sample, return others as-is.
         return output_class(
             {
-                k: (self.apply_mask(v, self._t_for_field(k, t)) if are_like_tensors(prev_sample, v) else v)
+                k: self.apply_mask(v, self._t_for_field(k, t)) if are_like_tensors(prev_sample, v) else v
                 for k, v in step_output.items()
             }
         )
@@ -162,39 +165,23 @@ class ControlNetData:
 
 
 @dataclass
-class ConditioningData:
-    unconditioned_embeddings: BasicConditioningInfo
-    text_embeddings: BasicConditioningInfo
-    guidance_scale: Union[float, List[float]]
-    """
-    Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
-    `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
-    Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate
-    images that are closely linked to the text `prompt`, usually at the expense of lower image quality.
-    """
-    extra: Optional[InvokeAIDiffuserComponent.ExtraConditioningInfo] = None
-    scheduler_args: dict[str, Any] = field(default_factory=dict)
-    """
-    Additional arguments to pass to invokeai_diffuser.do_latent_postprocessing().
-    """
-    postprocessing_settings: Optional[PostprocessingSettings] = None
+class IPAdapterData:
+    ip_adapter_model: IPAdapter = Field(default=None)
+    # TODO: change to polymorphic so can do different weights per step (once implemented...)
+    weight: Union[float, List[float]] = Field(default=1.0)
+    # weight: float = Field(default=1.0)
+    begin_step_percent: float = Field(default=0.0)
+    end_step_percent: float = Field(default=1.0)
 
-    @property
-    def dtype(self):
-        return self.text_embeddings.dtype
 
-    def add_scheduler_args_if_applicable(self, scheduler, **kwargs):
-        scheduler_args = dict(self.scheduler_args)
-        step_method = inspect.signature(scheduler.step)
-        for name, value in kwargs.items():
-            try:
-                step_method.bind_partial(**{name: value})
-            except TypeError:
-                # FIXME: don't silently discard arguments
-                pass  # debug("%s does not accept argument named %r", scheduler, name)
-            else:
-                scheduler_args[name] = value
-        return dataclasses.replace(self, scheduler_args=scheduler_args)
+@dataclass
+class T2IAdapterData:
+    """A structure containing the information required to apply conditioning from a single T2I-Adapter model."""
+
+    adapter_state: dict[torch.Tensor] = Field()
+    weight: Union[float, list[float]] = Field(default=1.0)
+    begin_step_percent: float = Field(default=0.0)
+    end_step_percent: float = Field(default=1.0)
 
 
 @dataclass
@@ -277,6 +264,7 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         )
         self.invokeai_diffuser = InvokeAIDiffuserComponent(self.unet, self._unet_forward)
         self.control_model = control_model
+        self.use_ip_adapter = False
 
     def _adjust_memory_efficient_attention(self, latents: torch.Tensor):
         """
@@ -349,6 +337,8 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         additional_guidance: List[Callable] = None,
         callback: Callable[[PipelineIntermediateState], None] = None,
         control_data: List[ControlNetData] = None,
+        ip_adapter_data: Optional[list[IPAdapterData]] = None,
+        t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
         mask: Optional[torch.Tensor] = None,
         masked_latents: Optional[torch.Tensor] = None,
         seed: Optional[int] = None,
@@ -400,6 +390,8 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                 conditioning_data,
                 additional_guidance=additional_guidance,
                 control_data=control_data,
+                ip_adapter_data=ip_adapter_data,
+                t2i_adapter_data=t2i_adapter_data,
                 callback=callback,
             )
         finally:
@@ -419,6 +411,8 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         *,
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
+        ip_adapter_data: Optional[list[IPAdapterData]] = None,
+        t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
         callback: Callable[[PipelineIntermediateState], None] = None,
     ):
         self._adjust_memory_efficient_attention(latents)
@@ -431,12 +425,24 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         if timesteps.shape[0] == 0:
             return latents, attention_map_saver
 
-        extra_conditioning_info = conditioning_data.extra
-        with self.invokeai_diffuser.custom_attention_context(
-            self.invokeai_diffuser.model,
-            extra_conditioning_info=extra_conditioning_info,
-            step_count=len(self.scheduler.timesteps),
-        ):
+        ip_adapter_unet_patcher = None
+        if conditioning_data.extra is not None and conditioning_data.extra.wants_cross_attention_control:
+            attn_ctx = self.invokeai_diffuser.custom_attention_context(
+                self.invokeai_diffuser.model,
+                extra_conditioning_info=conditioning_data.extra,
+                step_count=len(self.scheduler.timesteps),
+            )
+            self.use_ip_adapter = False
+        elif ip_adapter_data is not None:
+            # TODO(ryand): Should we raise an exception if both custom attention and IP-Adapter attention are active?
+            # As it is now, the IP-Adapter will silently be skipped.
+            ip_adapter_unet_patcher = UNetPatcher([ipa.ip_adapter_model for ipa in ip_adapter_data])
+            attn_ctx = ip_adapter_unet_patcher.apply_ip_adapter_attention(self.invokeai_diffuser.model)
+            self.use_ip_adapter = True
+        else:
+            attn_ctx = nullcontext()
+
+        with attn_ctx:
             if callback is not None:
                 callback(
                     PipelineIntermediateState(
@@ -459,6 +465,9 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                     total_step_count=len(timesteps),
                     additional_guidance=additional_guidance,
                     control_data=control_data,
+                    ip_adapter_data=ip_adapter_data,
+                    t2i_adapter_data=t2i_adapter_data,
+                    ip_adapter_unet_patcher=ip_adapter_unet_patcher,
                 )
                 latents = step_output.prev_sample
 
@@ -504,6 +513,9 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         total_step_count: int,
         additional_guidance: List[Callable] = None,
         control_data: List[ControlNetData] = None,
+        ip_adapter_data: Optional[list[IPAdapterData]] = None,
+        t2i_adapter_data: Optional[list[T2IAdapterData]] = None,
+        ip_adapter_unet_patcher: Optional[UNetPatcher] = None,
     ):
         # invokeai_diffuser has batched timesteps, but diffusers schedulers expect a single value
         timestep = t[0]
@@ -514,10 +526,32 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         #     i.e. before or after passing it to InvokeAIDiffuserComponent
         latent_model_input = self.scheduler.scale_model_input(latents, timestep)
 
-        # default is no controlnet, so set controlnet processing output to None
-        controlnet_down_block_samples, controlnet_mid_block_sample = None, None
-        if control_data is not None:
-            controlnet_down_block_samples, controlnet_mid_block_sample = self.invokeai_diffuser.do_controlnet_step(
+        # handle IP-Adapter
+        if self.use_ip_adapter and ip_adapter_data is not None:  # somewhat redundant but logic is clearer
+            for i, single_ip_adapter_data in enumerate(ip_adapter_data):
+                first_adapter_step = math.floor(single_ip_adapter_data.begin_step_percent * total_step_count)
+                last_adapter_step = math.ceil(single_ip_adapter_data.end_step_percent * total_step_count)
+                weight = (
+                    single_ip_adapter_data.weight[step_index]
+                    if isinstance(single_ip_adapter_data.weight, List)
+                    else single_ip_adapter_data.weight
+                )
+                if step_index >= first_adapter_step and step_index <= last_adapter_step:
+                    # Only apply this IP-Adapter if the current step is within the IP-Adapter's begin/end step range.
+                    ip_adapter_unet_patcher.set_scale(i, weight)
+                else:
+                    # Otherwise, set the IP-Adapter's scale to 0, so it has no effect.
+                    ip_adapter_unet_patcher.set_scale(i, 0.0)
+
+        # Handle ControlNet(s) and T2I-Adapter(s)
+        down_block_additional_residuals = None
+        mid_block_additional_residual = None
+        if control_data is not None and t2i_adapter_data is not None:
+            # TODO(ryand): This is a limitation of the UNet2DConditionModel API, not a fundamental incompatibility
+            # between ControlNets and T2I-Adapters. We will try to fix this upstream in diffusers.
+            raise Exception("ControlNet(s) and T2I-Adapter(s) cannot be used simultaneously (yet).")
+        elif control_data is not None:
+            down_block_additional_residuals, mid_block_additional_residual = self.invokeai_diffuser.do_controlnet_step(
                 control_data=control_data,
                 sample=latent_model_input,
                 timestep=timestep,
@@ -525,6 +559,32 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
                 total_step_count=total_step_count,
                 conditioning_data=conditioning_data,
             )
+        elif t2i_adapter_data is not None:
+            accum_adapter_state = None
+            for single_t2i_adapter_data in t2i_adapter_data:
+                # Determine the T2I-Adapter weights for the current denoising step.
+                first_t2i_adapter_step = math.floor(single_t2i_adapter_data.begin_step_percent * total_step_count)
+                last_t2i_adapter_step = math.ceil(single_t2i_adapter_data.end_step_percent * total_step_count)
+                t2i_adapter_weight = (
+                    single_t2i_adapter_data.weight[step_index]
+                    if isinstance(single_t2i_adapter_data.weight, list)
+                    else single_t2i_adapter_data.weight
+                )
+                if step_index < first_t2i_adapter_step or step_index > last_t2i_adapter_step:
+                    # If the current step is outside of the T2I-Adapter's begin/end step range, then set its weight to 0
+                    # so it has no effect.
+                    t2i_adapter_weight = 0.0
+
+                # Apply the t2i_adapter_weight, and accumulate.
+                if accum_adapter_state is None:
+                    # Handle the first T2I-Adapter.
+                    accum_adapter_state = [val * t2i_adapter_weight for val in single_t2i_adapter_data.adapter_state]
+                else:
+                    # Add to the previous adapter states.
+                    for idx, value in enumerate(single_t2i_adapter_data.adapter_state):
+                        accum_adapter_state[idx] += value * t2i_adapter_weight
+
+            down_block_additional_residuals = accum_adapter_state
 
         uc_noise_pred, c_noise_pred = self.invokeai_diffuser.do_unet_step(
             sample=latent_model_input,
@@ -533,8 +593,8 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             total_step_count=total_step_count,
             conditioning_data=conditioning_data,
             # extra:
-            down_block_additional_residuals=controlnet_down_block_samples,  # from controlnet(s)
-            mid_block_additional_residual=controlnet_mid_block_sample,  # from controlnet(s)
+            down_block_additional_residuals=down_block_additional_residuals,
+            mid_block_additional_residual=mid_block_additional_residual,
         )
 
         guidance_scale = conditioning_data.guidance_scale

invokeai/backend/stable_diffusion/diffusion/__init__.py

@@ -3,9 +3,4 @@ Initialization file for invokeai.models.diffusion
 """
 from .cross_attention_control import InvokeAICrossAttentionMixin  # noqa: F401
 from .cross_attention_map_saving import AttentionMapSaver  # noqa: F401
-from .shared_invokeai_diffusion import (  # noqa: F401
-    BasicConditioningInfo,
-    InvokeAIDiffuserComponent,
-    PostprocessingSettings,
-    SDXLConditioningInfo,
-)
+from .shared_invokeai_diffusion import InvokeAIDiffuserComponent  # noqa: F401

invokeai/backend/stable_diffusion/diffusion/conditioning_data.py

@@ -0,0 +1,101 @@
+import dataclasses
+import inspect
+from dataclasses import dataclass, field
+from typing import Any, List, Optional, Union
+
+import torch
+
+from .cross_attention_control import Arguments
+
+
+@dataclass
+class ExtraConditioningInfo:
+    tokens_count_including_eos_bos: int
+    cross_attention_control_args: Optional[Arguments] = None
+
+    @property
+    def wants_cross_attention_control(self):
+        return self.cross_attention_control_args is not None
+
+
+@dataclass
+class BasicConditioningInfo:
+    embeds: torch.Tensor
+    # TODO(ryand): Right now we awkwardly copy the extra conditioning info from here up to `ConditioningData`. This
+    # should only be stored in one place.
+    extra_conditioning: Optional[ExtraConditioningInfo]
+    # weight: float
+    # mode: ConditioningAlgo
+
+    def to(self, device, dtype=None):
+        self.embeds = self.embeds.to(device=device, dtype=dtype)
+        return self
+
+
+@dataclass
+class SDXLConditioningInfo(BasicConditioningInfo):
+    pooled_embeds: torch.Tensor
+    add_time_ids: torch.Tensor
+
+    def to(self, device, dtype=None):
+        self.pooled_embeds = self.pooled_embeds.to(device=device, dtype=dtype)
+        self.add_time_ids = self.add_time_ids.to(device=device, dtype=dtype)
+        return super().to(device=device, dtype=dtype)
+
+
+@dataclass(frozen=True)
+class PostprocessingSettings:
+    threshold: float
+    warmup: float
+    h_symmetry_time_pct: Optional[float]
+    v_symmetry_time_pct: Optional[float]
+
+
+@dataclass
+class IPAdapterConditioningInfo:
+    cond_image_prompt_embeds: torch.Tensor
+    """IP-Adapter image encoder conditioning embeddings.
+    Shape: (batch_size, num_tokens, encoding_dim).
+    """
+    uncond_image_prompt_embeds: torch.Tensor
+    """IP-Adapter image encoding embeddings to use for unconditional generation.
+    Shape: (batch_size, num_tokens, encoding_dim).
+    """
+
+
+@dataclass
+class ConditioningData:
+    unconditioned_embeddings: BasicConditioningInfo
+    text_embeddings: BasicConditioningInfo
+    guidance_scale: Union[float, List[float]]
+    """
+    Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+    `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf).
+    Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate
+    images that are closely linked to the text `prompt`, usually at the expense of lower image quality.
+    """
+    extra: Optional[ExtraConditioningInfo] = None
+    scheduler_args: dict[str, Any] = field(default_factory=dict)
+    """
+    Additional arguments to pass to invokeai_diffuser.do_latent_postprocessing().
+    """
+    postprocessing_settings: Optional[PostprocessingSettings] = None
+
+    ip_adapter_conditioning: Optional[list[IPAdapterConditioningInfo]] = None
+
+    @property
+    def dtype(self):
+        return self.text_embeddings.dtype
+
+    def add_scheduler_args_if_applicable(self, scheduler, **kwargs):
+        scheduler_args = dict(self.scheduler_args)
+        step_method = inspect.signature(scheduler.step)
+        for name, value in kwargs.items():
+            try:
+                step_method.bind_partial(**{name: value})
+            except TypeError:
+                # FIXME: don't silently discard arguments
+                pass  # debug("%s does not accept argument named %r", scheduler, name)
+            else:
+                scheduler_args[name] = value
+        return dataclasses.replace(self, scheduler_args=scheduler_args)

invokeai/backend/stable_diffusion/diffusion/cross_attention_control.py

@@ -376,11 +376,11 @@ def get_cross_attention_modules(model, which: CrossAttentionType) -> list[tuple[
         # non-fatal error but .swap() won't work.
         logger.error(
             f"Error! CrossAttentionControl found an unexpected number of {cross_attention_class} modules in the model "
-            + f"(expected {expected_count}, found {cross_attention_modules_in_model_count}). Either monkey-patching failed "
-            + "or some assumption has changed about the structure of the model itself. Please fix the monkey-patching, "
-            + f"and/or update the {expected_count} above to an appropriate number, and/or find and inform someone who knows "
-            + "what it means. This error is non-fatal, but it is likely that .swap() and attention map display will not "
-            + "work properly until it is fixed."
+            f"(expected {expected_count}, found {cross_attention_modules_in_model_count}). Either monkey-patching "
+            "failed or some assumption has changed about the structure of the model itself. Please fix the "
+            f"monkey-patching, and/or update the {expected_count} above to an appropriate number, and/or find and "
+            "inform someone who knows what it means. This error is non-fatal, but it is likely that .swap() and "
+            "attention map display will not work properly until it is fixed."
         )
     return attention_module_tuples
 
@@ -577,6 +577,7 @@ class SlicedSwapCrossAttnProcesser(SlicedAttnProcessor):
         attention_mask=None,
         # kwargs
         swap_cross_attn_context: SwapCrossAttnContext = None,
+        **kwargs,
     ):
         attention_type = CrossAttentionType.SELF if encoder_hidden_states is None else CrossAttentionType.TOKENS
 

invokeai/backend/stable_diffusion/diffusion/shared_invokeai_diffusion.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import math
 from contextlib import contextmanager
-from dataclasses import dataclass
 from typing import Any, Callable, Optional, Union
 
 import torch
@@ -10,9 +9,14 @@ from diffusers import UNet2DConditionModel
 from typing_extensions import TypeAlias
 
 from invokeai.app.services.config import InvokeAIAppConfig
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
+    ConditioningData,
+    ExtraConditioningInfo,
+    PostprocessingSettings,
+    SDXLConditioningInfo,
+)
 
 from .cross_attention_control import (
-    Arguments,
     Context,
     CrossAttentionType,
     SwapCrossAttnContext,
@@ -31,37 +35,6 @@ ModelForwardCallback: TypeAlias = Union[
 ]
 
 
-@dataclass
-class BasicConditioningInfo:
-    embeds: torch.Tensor
-    extra_conditioning: Optional[InvokeAIDiffuserComponent.ExtraConditioningInfo]
-    # weight: float
-    # mode: ConditioningAlgo
-
-    def to(self, device, dtype=None):
-        self.embeds = self.embeds.to(device=device, dtype=dtype)
-        return self
-
-
-@dataclass
-class SDXLConditioningInfo(BasicConditioningInfo):
-    pooled_embeds: torch.Tensor
-    add_time_ids: torch.Tensor
-
-    def to(self, device, dtype=None):
-        self.pooled_embeds = self.pooled_embeds.to(device=device, dtype=dtype)
-        self.add_time_ids = self.add_time_ids.to(device=device, dtype=dtype)
-        return super().to(device=device, dtype=dtype)
-
-
-@dataclass(frozen=True)
-class PostprocessingSettings:
-    threshold: float
-    warmup: float
-    h_symmetry_time_pct: Optional[float]
-    v_symmetry_time_pct: Optional[float]
-
-
 class InvokeAIDiffuserComponent:
     """
     The aim of this component is to provide a single place for code that can be applied identically to
@@ -75,15 +48,6 @@ class InvokeAIDiffuserComponent:
     debug_thresholding = False
     sequential_guidance = False
 
-    @dataclass
-    class ExtraConditioningInfo:
-        tokens_count_including_eos_bos: int
-        cross_attention_control_args: Optional[Arguments] = None
-
-        @property
-        def wants_cross_attention_control(self):
-            return self.cross_attention_control_args is not None
-
     def __init__(
         self,
         model,
@@ -103,30 +67,26 @@ class InvokeAIDiffuserComponent:
     @contextmanager
     def custom_attention_context(
         self,
-        unet: UNet2DConditionModel,  # note: also may futz with the text encoder depending on requested LoRAs
+        unet: UNet2DConditionModel,
         extra_conditioning_info: Optional[ExtraConditioningInfo],
         step_count: int,
     ):
-        old_attn_processors = None
-        if extra_conditioning_info and (extra_conditioning_info.wants_cross_attention_control):
-            old_attn_processors = unet.attn_processors
-            # Load lora conditions into the model
-            if extra_conditioning_info.wants_cross_attention_control:
-                self.cross_attention_control_context = Context(
-                    arguments=extra_conditioning_info.cross_attention_control_args,
-                    step_count=step_count,
-                )
-                setup_cross_attention_control_attention_processors(
-                    unet,
-                    self.cross_attention_control_context,
-                )
+        old_attn_processors = unet.attn_processors
 
         try:
+            self.cross_attention_control_context = Context(
+                arguments=extra_conditioning_info.cross_attention_control_args,
+                step_count=step_count,
+            )
+            setup_cross_attention_control_attention_processors(
+                unet,
+                self.cross_attention_control_context,
+            )
+
             yield None
         finally:
             self.cross_attention_control_context = None
-            if old_attn_processors is not None:
-                unet.set_attn_processor(old_attn_processors)
+            unet.set_attn_processor(old_attn_processors)
             # TODO resuscitate attention map saving
             # self.remove_attention_map_saving()
 
@@ -376,11 +336,22 @@ class InvokeAIDiffuserComponent:
 
     # methods below are called from do_diffusion_step and should be considered private to this class.
 
-    def _apply_standard_conditioning(self, x, sigma, conditioning_data, **kwargs):
-        # fast batched path
+    def _apply_standard_conditioning(self, x, sigma, conditioning_data: ConditioningData, **kwargs):
+        """Runs the conditioned and unconditioned UNet forward passes in a single batch for faster inference speed at
+        the cost of higher memory usage.
+        """
         x_twice = torch.cat([x] * 2)
         sigma_twice = torch.cat([sigma] * 2)
 
+        cross_attention_kwargs = None
+        if conditioning_data.ip_adapter_conditioning is not None:
+            cross_attention_kwargs = {
+                "ip_adapter_image_prompt_embeds": [
+                    torch.cat([ipa_conditioning.uncond_image_prompt_embeds, ipa_conditioning.cond_image_prompt_embeds])
+                    for ipa_conditioning in conditioning_data.ip_adapter_conditioning
+                ]
+            }
+
         added_cond_kwargs = None
         if type(conditioning_data.text_embeddings) is SDXLConditioningInfo:
             added_cond_kwargs = {
@@ -408,6 +379,7 @@ class InvokeAIDiffuserComponent:
             x_twice,
             sigma_twice,
             both_conditionings,
+            cross_attention_kwargs=cross_attention_kwargs,
             encoder_attention_mask=encoder_attention_mask,
             added_cond_kwargs=added_cond_kwargs,
             **kwargs,
@@ -419,9 +391,12 @@ class InvokeAIDiffuserComponent:
         self,
         x: torch.Tensor,
         sigma,
-        conditioning_data,
+        conditioning_data: ConditioningData,
         **kwargs,
     ):
+        """Runs the conditioned and unconditioned UNet forward passes sequentially for lower memory usage at the cost of
+        slower execution speed.
+        """
         # low-memory sequential path
         uncond_down_block, cond_down_block = None, None
         down_block_additional_residuals = kwargs.pop("down_block_additional_residuals", None)
@@ -437,6 +412,16 @@ class InvokeAIDiffuserComponent:
         if mid_block_additional_residual is not None:
             uncond_mid_block, cond_mid_block = mid_block_additional_residual.chunk(2)
 
+        # Run unconditional UNet denoising.
+        cross_attention_kwargs = None
+        if conditioning_data.ip_adapter_conditioning is not None:
+            cross_attention_kwargs = {
+                "ip_adapter_image_prompt_embeds": [
+                    ipa_conditioning.uncond_image_prompt_embeds
+                    for ipa_conditioning in conditioning_data.ip_adapter_conditioning
+                ]
+            }
+
         added_cond_kwargs = None
         is_sdxl = type(conditioning_data.text_embeddings) is SDXLConditioningInfo
         if is_sdxl:
@@ -449,12 +434,24 @@ class InvokeAIDiffuserComponent:
             x,
             sigma,
             conditioning_data.unconditioned_embeddings.embeds,
+            cross_attention_kwargs=cross_attention_kwargs,
             down_block_additional_residuals=uncond_down_block,
             mid_block_additional_residual=uncond_mid_block,
             added_cond_kwargs=added_cond_kwargs,
             **kwargs,
         )
 
+        # Run conditional UNet denoising.
+        cross_attention_kwargs = None
+        if conditioning_data.ip_adapter_conditioning is not None:
+            cross_attention_kwargs = {
+                "ip_adapter_image_prompt_embeds": [
+                    ipa_conditioning.cond_image_prompt_embeds
+                    for ipa_conditioning in conditioning_data.ip_adapter_conditioning
+                ]
+            }
+
+        added_cond_kwargs = None
         if is_sdxl:
             added_cond_kwargs = {
                 "text_embeds": conditioning_data.text_embeddings.pooled_embeds,
@@ -465,6 +462,7 @@ class InvokeAIDiffuserComponent:
             x,
             sigma,
             conditioning_data.text_embeddings.embeds,
+            cross_attention_kwargs=cross_attention_kwargs,
             down_block_additional_residuals=cond_down_block,
             mid_block_additional_residual=cond_mid_block,
             added_cond_kwargs=added_cond_kwargs,

invokeai/backend/util/db_maintenance.py

@@ -0,0 +1,568 @@
+# pylint: disable=line-too-long
+# pylint: disable=broad-exception-caught
+# pylint: disable=missing-function-docstring
+"""Script to peform db maintenance and outputs directory management."""
+
+import argparse
+import datetime
+import enum
+import glob
+import locale
+import os
+import shutil
+import sqlite3
+from pathlib import Path
+
+import PIL
+import PIL.ImageOps
+import PIL.PngImagePlugin
+import yaml
+
+
+class ConfigMapper:
+    """Configuration loader."""
+
+    def __init__(self):  # noqa D107
+        pass
+
+    TIMESTAMP_STRING = datetime.datetime.utcnow().strftime("%Y%m%dT%H%M%SZ")
+
+    INVOKE_DIRNAME = "invokeai"
+    YAML_FILENAME = "invokeai.yaml"
+    DATABASE_FILENAME = "invokeai.db"
+
+    database_path = None
+    database_backup_dir = None
+    outputs_path = None
+    archive_path = None
+    thumbnails_path = None
+    thumbnails_archive_path = None
+
+    def load(self):
+        """Read paths from yaml config and validate."""
+        root = "."
+
+        if not self.__load_from_root_config(os.path.abspath(root)):
+            return False
+
+        return True
+
+    def __load_from_root_config(self, invoke_root):
+        """Validate a yaml path exists, confirm the user wants to use it and load config."""
+        yaml_path = os.path.join(invoke_root, self.YAML_FILENAME)
+        if os.path.exists(yaml_path):
+            db_dir, outdir = self.__load_paths_from_yaml_file(yaml_path)
+
+            if db_dir is None or outdir is None:
+                print("The invokeai.yaml file was found but is missing the db_dir and/or outdir setting!")
+                return False
+
+            if os.path.isabs(db_dir):
+                self.database_path = os.path.join(db_dir, self.DATABASE_FILENAME)
+            else:
+                self.database_path = os.path.join(invoke_root, db_dir, self.DATABASE_FILENAME)
+
+            self.database_backup_dir = os.path.join(os.path.dirname(self.database_path), "backup")
+
+            if os.path.isabs(outdir):
+                self.outputs_path = os.path.join(outdir, "images")
+                self.archive_path = os.path.join(outdir, "images-archive")
+            else:
+                self.outputs_path = os.path.join(invoke_root, outdir, "images")
+                self.archive_path = os.path.join(invoke_root, outdir, "images-archive")
+
+            self.thumbnails_path = os.path.join(self.outputs_path, "thumbnails")
+            self.thumbnails_archive_path = os.path.join(self.archive_path, "thumbnails")
+
+            db_exists = os.path.exists(self.database_path)
+            outdir_exists = os.path.exists(self.outputs_path)
+
+            text = f"Found {self.YAML_FILENAME} file at {yaml_path}:"
+            text += f"\n  Database : {self.database_path} - {'Exists!' if db_exists else 'Not Found!'}"
+            text += f"\n  Outputs  : {self.outputs_path}- {'Exists!' if outdir_exists else 'Not Found!'}"
+            print(text)
+
+            if db_exists and outdir_exists:
+                return True
+            else:
+                print(
+                    "\nOne or more paths specified in invoke.yaml do not exist. Please inspect/correct the configuration and ensure the script is run in the developer console mode (option 8) from an Invoke AI root directory."
+                )
+                return False
+        else:
+            print(
+                f"Auto-discovery of configuration failed! Could not find ({yaml_path})!\n\nPlease ensure the script is run in the developer console mode (option 8) from an Invoke AI root directory."
+            )
+            return False
+
+    def __load_paths_from_yaml_file(self, yaml_path):
+        """Load an Invoke AI yaml file and get the database and outputs paths."""
+        try:
+            with open(yaml_path, "rt", encoding=locale.getpreferredencoding()) as file:
+                yamlinfo = yaml.safe_load(file)
+                db_dir = yamlinfo.get("InvokeAI", {}).get("Paths", {}).get("db_dir", None)
+                outdir = yamlinfo.get("InvokeAI", {}).get("Paths", {}).get("outdir", None)
+                return db_dir, outdir
+        except Exception:
+            print(f"Failed to load paths from yaml file! {yaml_path}!")
+            return None, None
+
+
+class MaintenanceStats:
+    """DTO for tracking work progress."""
+
+    def __init__(self):  # noqa D107
+        pass
+
+    time_start = datetime.datetime.utcnow()
+    count_orphaned_db_entries_cleaned = 0
+    count_orphaned_disk_files_cleaned = 0
+    count_orphaned_thumbnails_cleaned = 0
+    count_thumbnails_regenerated = 0
+    count_errors = 0
+
+    @staticmethod
+    def get_elapsed_time_string():
+        """Get a friendly time string for the  time elapsed since processing start."""
+        time_now = datetime.datetime.utcnow()
+        total_seconds = (time_now - MaintenanceStats.time_start).total_seconds()
+        hours = int((total_seconds) / 3600)
+        minutes = int(((total_seconds) % 3600) / 60)
+        seconds = total_seconds % 60
+        out_str = f"{hours} hour(s) -" if hours > 0 else ""
+        out_str += f"{minutes} minute(s) -" if minutes > 0 else ""
+        out_str += f"{seconds:.2f} second(s)"
+        return out_str
+
+
+class DatabaseMapper:
+    """Class to abstract database functionality."""
+
+    def __init__(self, database_path, database_backup_dir):  # noqa D107
+        self.database_path = database_path
+        self.database_backup_dir = database_backup_dir
+        self.connection = None
+        self.cursor = None
+
+    def backup(self, timestamp_string):
+        """Take a backup of the database."""
+        if not os.path.exists(self.database_backup_dir):
+            print(f"Database backup directory {self.database_backup_dir} does not exist -> creating...", end="")
+            os.makedirs(self.database_backup_dir)
+            print("Done!")
+        database_backup_path = os.path.join(self.database_backup_dir, f"backup-{timestamp_string}-invokeai.db")
+        print(f"Making DB Backup at {database_backup_path}...", end="")
+        shutil.copy2(self.database_path, database_backup_path)
+        print("Done!")
+
+    def connect(self):
+        """Open connection to the database."""
+        self.connection = sqlite3.connect(self.database_path)
+        self.cursor = self.connection.cursor()
+
+    def get_all_image_files(self):
+        """Get the full list of image file names from the database."""
+        sql_get_image_by_name = "SELECT image_name FROM images"
+        self.cursor.execute(sql_get_image_by_name)
+        rows = self.cursor.fetchall()
+        db_files = []
+        for row in rows:
+            db_files.append(row[0])
+        return db_files
+
+    def remove_image_file_record(self, filename: str):
+        """Remove an image file reference from the database by filename."""
+        sanitized_filename = str.replace(filename, "'", "''")  # prevent injection
+        sql_command = f"DELETE FROM images WHERE image_name='{sanitized_filename}'"
+        self.cursor.execute(sql_command)
+        self.connection.commit()
+
+    def does_image_exist(self, image_filename):
+        """Check database if a image name already exists and return a boolean."""
+        sanitized_filename = str.replace(image_filename, "'", "''")  # prevent injection
+        sql_get_image_by_name = f"SELECT image_name FROM images WHERE image_name='{sanitized_filename}'"
+        self.cursor.execute(sql_get_image_by_name)
+        rows = self.cursor.fetchall()
+        return True if len(rows) > 0 else False
+
+    def disconnect(self):
+        """Disconnect from the db, cleaning up connections and cursors."""
+        if self.cursor is not None:
+            self.cursor.close()
+        if self.connection is not None:
+            self.connection.close()
+
+
+class PhysicalFileMapper:
+    """Containing class for script functionality."""
+
+    def __init__(self, outputs_path, thumbnails_path, archive_path, thumbnails_archive_path):  # noqa D107
+        self.outputs_path = outputs_path
+        self.archive_path = archive_path
+        self.thumbnails_path = thumbnails_path
+        self.thumbnails_archive_path = thumbnails_archive_path
+
+    def create_archive_directories(self):
+        """Create the directory for archiving orphaned image files."""
+        if not os.path.exists(self.archive_path):
+            print(f"Image archive directory ({self.archive_path}) does not exist -> creating...", end="")
+            os.makedirs(self.archive_path)
+            print("Created!")
+        if not os.path.exists(self.thumbnails_archive_path):
+            print(
+                f"Image thumbnails archive directory ({self.thumbnails_archive_path}) does not exist -> creating...",
+                end="",
+            )
+            os.makedirs(self.thumbnails_archive_path)
+            print("Created!")
+
+    def get_image_path_for_image_name(self, image_filename):  # noqa D102
+        return os.path.join(self.outputs_path, image_filename)
+
+    def image_file_exists(self, image_filename):  # noqa D102
+        return os.path.exists(self.get_image_path_for_image_name(image_filename))
+
+    def get_thumbnail_path_for_image(self, image_filename):  # noqa D102
+        return os.path.join(self.thumbnails_path, os.path.splitext(image_filename)[0]) + ".webp"
+
+    def get_image_name_from_thumbnail_path(self, thumbnail_path):  # noqa D102
+        return os.path.splitext(os.path.basename(thumbnail_path))[0] + ".png"
+
+    def thumbnail_exists_for_filename(self, image_filename):  # noqa D102
+        return os.path.exists(self.get_thumbnail_path_for_image(image_filename))
+
+    def archive_image(self, image_filename):  # noqa D102
+        if self.image_file_exists(image_filename):
+            image_path = self.get_image_path_for_image_name(image_filename)
+            shutil.move(image_path, self.archive_path)
+
+    def archive_thumbnail_by_image_filename(self, image_filename):  # noqa D102
+        if self.thumbnail_exists_for_filename(image_filename):
+            thumbnail_path = self.get_thumbnail_path_for_image(image_filename)
+            shutil.move(thumbnail_path, self.thumbnails_archive_path)
+
+    def get_all_png_filenames_in_directory(self, directory_path):  # noqa D102
+        filepaths = glob.glob(directory_path + "/*.png", recursive=False)
+        filenames = []
+        for filepath in filepaths:
+            filenames.append(os.path.basename(filepath))
+        return filenames
+
+    def get_all_thumbnails_with_full_path(self, thumbnails_directory):  # noqa D102
+        return glob.glob(thumbnails_directory + "/*.webp", recursive=False)
+
+    def generate_thumbnail_for_image_name(self, image_filename):  # noqa D102
+        # create thumbnail
+        file_path = self.get_image_path_for_image_name(image_filename)
+        thumb_path = self.get_thumbnail_path_for_image(image_filename)
+        thumb_size = 256, 256
+        with PIL.Image.open(file_path) as source_image:
+            source_image.thumbnail(thumb_size)
+            source_image.save(thumb_path, "webp")
+
+
+class MaintenanceOperation(str, enum.Enum):
+    """Enum class for operations."""
+
+    Ask = "ask"
+    CleanOrphanedDbEntries = "clean"
+    CleanOrphanedDiskFiles = "archive"
+    ReGenerateThumbnails = "thumbnails"
+    All = "all"
+
+
+class InvokeAIDatabaseMaintenanceApp:
+    """Main processor class for the application."""
+
+    _operation: MaintenanceOperation
+    _headless: bool = False
+    __stats: MaintenanceStats = MaintenanceStats()
+
+    def __init__(self, operation: MaintenanceOperation = MaintenanceOperation.Ask):
+        """Initialize maintenance app."""
+        self._operation = MaintenanceOperation(operation)
+        self._headless = operation != MaintenanceOperation.Ask
+
+    def ask_for_operation(self) -> MaintenanceOperation:
+        """Ask user to choose the operation to perform."""
+        while True:
+            print()
+            print("It is recommennded to run these operations as ordered below to avoid additional")
+            print("work being performed that will be discarded in a subsequent step.")
+            print()
+            print("Select maintenance operation:")
+            print()
+            print("1) Clean Orphaned Database Image Entries")
+            print("     Cleans entries in the database where the matching file was removed from")
+            print("     the outputs directory.")
+            print("2) Archive Orphaned Image Files")
+            print("     Files found in the outputs directory without an entry in the database are")
+            print("     moved to an archive directory.")
+            print("3) Re-Generate Missing Thumbnail Files")
+            print("     For files found in the outputs directory, re-generate a thumbnail if it")
+            print("     not found in the thumbnails directory.")
+            print()
+            print("(CTRL-C to quit)")
+
+            try:
+                input_option = int(input("Specify desired operation number (1-3): "))
+
+                operations = [
+                    MaintenanceOperation.CleanOrphanedDbEntries,
+                    MaintenanceOperation.CleanOrphanedDiskFiles,
+                    MaintenanceOperation.ReGenerateThumbnails,
+                ]
+                return operations[input_option - 1]
+            except (IndexError, ValueError):
+                print("\nInvalid selection!")
+
+    def ask_to_continue(self) -> bool:
+        """Ask user whether they want to continue with the operation."""
+        while True:
+            input_choice = input("Do you wish to continue? (Y or N)? ")
+            if str.lower(input_choice) == "y":
+                return True
+            if str.lower(input_choice) == "n":
+                return False
+
+    def clean_orphaned_db_entries(
+        self, config: ConfigMapper, file_mapper: PhysicalFileMapper, db_mapper: DatabaseMapper
+    ):
+        """Clean dangling database entries that no longer point to a file in outputs."""
+        if self._headless:
+            print(f"Removing database references to images that no longer exist in {config.outputs_path}...")
+        else:
+            print()
+            print("===============================================================================")
+            print("= Clean Orphaned Database Entries")
+            print()
+            print("Perform this operation if you have removed files from the outputs/images")
+            print("directory but the database was never updated. You may see this as empty imaages")
+            print("in the app gallery, or images that only show an enlarged version of the")
+            print("thumbnail.")
+            print()
+            print(f"Database File Path               : {config.database_path}")
+            print(f"Database backup will be taken at : {config.database_backup_dir}")
+            print(f"Outputs/Images Directory         : {config.outputs_path}")
+            print(f"Outputs/Images Archive Directory : {config.archive_path}")
+
+            print("\nNotes about this operation:")
+            print("- This operation will find database image file entries that do not exist in the")
+            print("  outputs/images dir and remove those entries from the database.")
+            print("- This operation will target all image types including intermediate files.")
+            print("- If a thumbnail still exists in outputs/images/thumbnails matching the")
+            print("  orphaned entry, it will be moved to the archive directory.")
+            print()
+
+            if not self.ask_to_continue():
+                raise KeyboardInterrupt
+
+        file_mapper.create_archive_directories()
+        db_mapper.backup(config.TIMESTAMP_STRING)
+        db_mapper.connect()
+        db_files = db_mapper.get_all_image_files()
+        for db_file in db_files:
+            try:
+                if not file_mapper.image_file_exists(db_file):
+                    print(f"Found orphaned image db entry {db_file}. Cleaning ...", end="")
+                    db_mapper.remove_image_file_record(db_file)
+                    print("Cleaned!")
+                    if file_mapper.thumbnail_exists_for_filename(db_file):
+                        print("A thumbnail was found, archiving ...", end="")
+                        file_mapper.archive_thumbnail_by_image_filename(db_file)
+                        print("Archived!")
+                    self.__stats.count_orphaned_db_entries_cleaned += 1
+            except Exception as ex:
+                print("An error occurred cleaning db entry, error was:")
+                print(ex)
+                self.__stats.count_errors += 1
+
+    def clean_orphaned_disk_files(
+        self, config: ConfigMapper, file_mapper: PhysicalFileMapper, db_mapper: DatabaseMapper
+    ):
+        """Archive image files that no longer have entries in the database."""
+        if self._headless:
+            print(f"Archiving orphaned image files to {config.archive_path}...")
+        else:
+            print()
+            print("===============================================================================")
+            print("= Clean Orphaned Disk Files")
+            print()
+            print("Perform this operation if you have files that were copied into the outputs")
+            print("directory which are not referenced by the database. This can happen if you")
+            print("upgraded to a version with a fresh database, but re-used the outputs directory")
+            print("and now new images are mixed with the files not in the db. The script will")
+            print("archive these files so you can choose to delete them or re-import using the")
+            print("official import script.")
+            print()
+            print(f"Database File Path               : {config.database_path}")
+            print(f"Database backup will be taken at : {config.database_backup_dir}")
+            print(f"Outputs/Images Directory         : {config.outputs_path}")
+            print(f"Outputs/Images Archive Directory : {config.archive_path}")
+
+            print("\nNotes about this operation:")
+            print("- This operation will find image files not referenced by the database and move to an")
+            print("  archive directory.")
+            print("- This operation will target all image types including intermediate references.")
+            print("- The matching thumbnail will also be archived.")
+            print("- Any remaining orphaned thumbnails will also be archived.")
+
+            if not self.ask_to_continue():
+                raise KeyboardInterrupt
+
+            print()
+
+        file_mapper.create_archive_directories()
+        db_mapper.backup(config.TIMESTAMP_STRING)
+        db_mapper.connect()
+        phys_files = file_mapper.get_all_png_filenames_in_directory(config.outputs_path)
+        for phys_file in phys_files:
+            try:
+                if not db_mapper.does_image_exist(phys_file):
+                    print(f"Found orphaned file {phys_file}, archiving...", end="")
+                    file_mapper.archive_image(phys_file)
+                    print("Archived!")
+                    if file_mapper.thumbnail_exists_for_filename(phys_file):
+                        print("Related thumbnail exists, archiving...", end="")
+                        file_mapper.archive_thumbnail_by_image_filename(phys_file)
+                        print("Archived!")
+                    else:
+                        print("No matching thumbnail existed to be cleaned.")
+                    self.__stats.count_orphaned_disk_files_cleaned += 1
+            except Exception as ex:
+                print("Error found trying to archive file or thumbnail, error was:")
+                print(ex)
+                self.__stats.count_errors += 1
+
+        thumb_filepaths = file_mapper.get_all_thumbnails_with_full_path(config.thumbnails_path)
+        # archive any remaining orphaned thumbnails
+        for thumb_filepath in thumb_filepaths:
+            try:
+                thumb_src_image_name = file_mapper.get_image_name_from_thumbnail_path(thumb_filepath)
+                if not file_mapper.image_file_exists(thumb_src_image_name):
+                    print(f"Found orphaned thumbnail {thumb_filepath}, archiving...", end="")
+                    file_mapper.archive_thumbnail_by_image_filename(thumb_src_image_name)
+                    print("Archived!")
+                    self.__stats.count_orphaned_thumbnails_cleaned += 1
+            except Exception as ex:
+                print("Error found trying to archive thumbnail, error was:")
+                print(ex)
+                self.__stats.count_errors += 1
+
+    def regenerate_thumbnails(self, config: ConfigMapper, file_mapper: PhysicalFileMapper, *args):
+        """Create missing thumbnails for any valid general images both in the db and on disk."""
+        if self._headless:
+            print("Regenerating missing image thumbnails...")
+        else:
+            print()
+            print("===============================================================================")
+            print("= Regenerate Thumbnails")
+            print()
+            print("This operation will find files that have no matching thumbnail on disk")
+            print("and regenerate those thumbnail files.")
+            print("NOTE: It is STRONGLY recommended that the user first clean/archive orphaned")
+            print("  disk files from the previous menu to avoid wasting time regenerating")
+            print("  thumbnails for orphaned files.")
+
+            print()
+            print(f"Outputs/Images Directory         : {config.outputs_path}")
+            print(f"Outputs/Images Directory         : {config.thumbnails_path}")
+
+            print("\nNotes about this operation:")
+            print("- This operation will find image files both referenced in the db and on disk")
+            print("  that do not have a matching thumbnail on disk and re-generate the thumbnail")
+            print("  file.")
+
+            if not self.ask_to_continue():
+                raise KeyboardInterrupt
+
+            print()
+
+        phys_files = file_mapper.get_all_png_filenames_in_directory(config.outputs_path)
+        for phys_file in phys_files:
+            try:
+                if not file_mapper.thumbnail_exists_for_filename(phys_file):
+                    print(f"Found file without thumbnail {phys_file}...Regenerating Thumbnail...", end="")
+                    file_mapper.generate_thumbnail_for_image_name(phys_file)
+                    print("Done!")
+                    self.__stats.count_thumbnails_regenerated += 1
+            except Exception as ex:
+                print("Error found trying to regenerate thumbnail, error was:")
+                print(ex)
+                self.__stats.count_errors += 1
+
+    def main(self):  # noqa D107
+        print("\n===============================================================================")
+        print("Database and outputs Maintenance for Invoke AI 3.0.0 +")
+        print("===============================================================================\n")
+
+        config_mapper = ConfigMapper()
+        if not config_mapper.load():
+            print("\nInvalid configuration...exiting.\n")
+            return
+
+        file_mapper = PhysicalFileMapper(
+            config_mapper.outputs_path,
+            config_mapper.thumbnails_path,
+            config_mapper.archive_path,
+            config_mapper.thumbnails_archive_path,
+        )
+        db_mapper = DatabaseMapper(config_mapper.database_path, config_mapper.database_backup_dir)
+
+        op = self._operation
+        operations_to_perform = []
+
+        if op == MaintenanceOperation.Ask:
+            op = self.ask_for_operation()
+
+        if op in [MaintenanceOperation.CleanOrphanedDbEntries, MaintenanceOperation.All]:
+            operations_to_perform.append(self.clean_orphaned_db_entries)
+        if op in [MaintenanceOperation.CleanOrphanedDiskFiles, MaintenanceOperation.All]:
+            operations_to_perform.append(self.clean_orphaned_disk_files)
+        if op in [MaintenanceOperation.ReGenerateThumbnails, MaintenanceOperation.All]:
+            operations_to_perform.append(self.regenerate_thumbnails)
+
+        for operation in operations_to_perform:
+            operation(config_mapper, file_mapper, db_mapper)
+
+        print("\n===============================================================================")
+        print(f"= Maintenance Complete - Elapsed Time: {MaintenanceStats.get_elapsed_time_string()}")
+        print()
+        print(f"Orphaned db entries cleaned             : {self.__stats.count_orphaned_db_entries_cleaned}")
+        print(f"Orphaned disk files archived            : {self.__stats.count_orphaned_disk_files_cleaned}")
+        print(f"Orphaned thumbnail files archived       : {self.__stats.count_orphaned_thumbnails_cleaned}")
+        print(f"Thumbnails regenerated                  : {self.__stats.count_thumbnails_regenerated}")
+        print(f"Errors during operation                 : {self.__stats.count_errors}")
+
+        print()
+
+
+def main():  # noqa D107
+    parser = argparse.ArgumentParser(
+        description="InvokeAI image database maintenance utility",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""Operations:
+  ask         Choose operation from a menu [default]
+  all         Run all maintenance operations
+  clean       Clean database of dangling entries
+  archive     Archive orphaned image files
+  thumbnails  Regenerate missing image thumbnails
+""",
+    )
+    parser.add_argument("--root", default=".", type=Path, help="InvokeAI root directory")
+    parser.add_argument(
+        "--operation", default="ask", choices=[x.value for x in MaintenanceOperation], help="Operation to perform."
+    )
+    args = parser.parse_args()
+    try:
+        os.chdir(args.root)
+        app = InvokeAIDatabaseMaintenanceApp(args.operation)
+        app.main()
+    except KeyboardInterrupt:
+        print("\n\nUser cancelled execution.")
+    except FileNotFoundError:
+        print(f"Invalid root directory '{args.root}'.")
+
+
+if __name__ == "__main__":
+    main()

invokeai/backend/util/hotfixes.py

@@ -24,7 +24,7 @@ from invokeai.backend.util.logging import InvokeAILogger
 # Modified ControlNetModel with encoder_attention_mask argument added
 
 
-logger = InvokeAILogger.getLogger(__name__)
+logger = InvokeAILogger.get_logger(__name__)
 
 
 class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlnetMixin):

invokeai/backend/util/logging.py

@@ -1,7 +1,6 @@
 # Copyright (c) 2023 Lincoln D. Stein and The InvokeAI Development Team
 
-"""
-invokeai.backend.util.logging
+"""invokeai.backend.util.logging
 
 Logging class for InvokeAI that produces console messages
 
@@ -9,9 +8,9 @@ Usage:
 
 from invokeai.backend.util.logging import InvokeAILogger
 
-logger = InvokeAILogger.getLogger(name='InvokeAI') // Initialization
+logger = InvokeAILogger.get_logger(name='InvokeAI') // Initialization
 (or)
-logger = InvokeAILogger.getLogger(__name__) // To use the filename
+logger = InvokeAILogger.get_logger(__name__) // To use the filename
 logger.configure()
 
 logger.critical('this is critical') // Critical Message
@@ -34,13 +33,13 @@ IAILogger.debug('this is a debugging message')
 ## Configuration
 
 The default configuration will print to stderr on the console. To add
-additional logging handlers, call getLogger with an initialized InvokeAIAppConfig
+additional logging handlers, call get_logger with an initialized InvokeAIAppConfig
 object:
 
 
  config = InvokeAIAppConfig.get_config()
  config.parse_args()
- logger = InvokeAILogger.getLogger(config=config)
+ logger = InvokeAILogger.get_logger(config=config)
 
 ### Three command-line options control logging:
 
@@ -173,6 +172,7 @@ InvokeAI:
     log_level: info
     log_format: color
 ```
+
 """
 
 import logging.handlers
@@ -193,39 +193,35 @@ except ImportError:
 
 # module level functions
 def debug(msg, *args, **kwargs):
-    InvokeAILogger.getLogger().debug(msg, *args, **kwargs)
+    InvokeAILogger.get_logger().debug(msg, *args, **kwargs)
 
 
 def info(msg, *args, **kwargs):
-    InvokeAILogger.getLogger().info(msg, *args, **kwargs)
+    InvokeAILogger.get_logger().info(msg, *args, **kwargs)
 
 
 def warning(msg, *args, **kwargs):
-    InvokeAILogger.getLogger().warning(msg, *args, **kwargs)
+    InvokeAILogger.get_logger().warning(msg, *args, **kwargs)
 
 
 def error(msg, *args, **kwargs):
-    InvokeAILogger.getLogger().error(msg, *args, **kwargs)
+    InvokeAILogger.get_logger().error(msg, *args, **kwargs)
 
 
 def critical(msg, *args, **kwargs):
-    InvokeAILogger.getLogger().critical(msg, *args, **kwargs)
+    InvokeAILogger.get_logger().critical(msg, *args, **kwargs)
 
 
 def log(level, msg, *args, **kwargs):
-    InvokeAILogger.getLogger().log(level, msg, *args, **kwargs)
+    InvokeAILogger.get_logger().log(level, msg, *args, **kwargs)
 
 
 def disable(level=logging.CRITICAL):
-    InvokeAILogger.getLogger().disable(level)
+    InvokeAILogger.get_logger().disable(level)
 
 
 def basicConfig(**kwargs):
-    InvokeAILogger.getLogger().basicConfig(**kwargs)
-
-
-def getLogger(name: str = None) -> logging.Logger:
-    return InvokeAILogger.getLogger(name)
+    InvokeAILogger.get_logger().basicConfig(**kwargs)
 
 
 _FACILITY_MAP = (
@@ -351,7 +347,7 @@ class InvokeAILogger(object):
     loggers = dict()
 
     @classmethod
-    def getLogger(
+    def get_logger(
         cls, name: str = "InvokeAI", config: InvokeAIAppConfig = InvokeAIAppConfig.get_config()
     ) -> logging.Logger:
         if name in cls.loggers:
@@ -360,13 +356,13 @@ class InvokeAILogger(object):
         else:
             logger = logging.getLogger(name)
         logger.setLevel(config.log_level.upper())  # yes, strings work here
-        for ch in cls.getLoggers(config):
+        for ch in cls.get_loggers(config):
             logger.addHandler(ch)
             cls.loggers[name] = logger
         return cls.loggers[name]
 
     @classmethod
-    def getLoggers(cls, config: InvokeAIAppConfig) -> list[logging.Handler]:
+    def get_loggers(cls, config: InvokeAIAppConfig) -> list[logging.Handler]:
         handler_strs = config.log_handlers
         handlers = list()
         for handler in handler_strs:

invokeai/backend/util/test_utils.py

@@ -0,0 +1,67 @@
+import contextlib
+from pathlib import Path
+from typing import Optional, Union
+
+import pytest
+import torch
+
+from invokeai.app.services.config.invokeai_config import InvokeAIAppConfig
+from invokeai.backend.install.model_install_backend import ModelInstall
+from invokeai.backend.model_management.model_manager import ModelInfo
+from invokeai.backend.model_management.models.base import BaseModelType, ModelNotFoundException, ModelType, SubModelType
+
+
+@pytest.fixture(scope="session")
+def torch_device():
+    return "cuda" if torch.cuda.is_available() else "cpu"
+
+
+@pytest.fixture(scope="module")
+def model_installer():
+    """A global ModelInstall pytest fixture to be used by many tests."""
+    # HACK(ryand): InvokeAIAppConfig.get_config() returns a singleton config object. This can lead to weird interactions
+    # between tests that need to alter the config. For example, some tests change the 'root' directory in the config,
+    # which can cause `install_and_load_model(...)` to re-download the model unnecessarily. As a temporary workaround,
+    # we pass a kwarg to get_config, which causes the config to be re-loaded. To fix this properly, we should stop using
+    # a singleton.
+    return ModelInstall(InvokeAIAppConfig.get_config(log_level="info"))
+
+
+def install_and_load_model(
+    model_installer: ModelInstall,
+    model_path_id_or_url: Union[str, Path],
+    model_name: str,
+    base_model: BaseModelType,
+    model_type: ModelType,
+    submodel_type: Optional[SubModelType] = None,
+) -> ModelInfo:
+    """Install a model if it is not already installed, then get the ModelInfo for that model.
+
+    This is intended as a utility function for tests.
+
+    Args:
+        model_installer (ModelInstall): The model installer.
+        model_path_id_or_url (Union[str, Path]): The path, HF ID, URL, etc. where the model can be installed from if it
+            is not already installed.
+        model_name (str): The model name, forwarded to ModelManager.get_model(...).
+        base_model (BaseModelType): The base model, forwarded to ModelManager.get_model(...).
+        model_type (ModelType): The model type, forwarded to ModelManager.get_model(...).
+        submodel_type (Optional[SubModelType]): The submodel type, forwarded to ModelManager.get_model(...).
+
+    Returns:
+        ModelInfo
+    """
+    # If the requested model is already installed, return its ModelInfo.
+    with contextlib.suppress(ModelNotFoundException):
+        return model_installer.mgr.get_model(model_name, base_model, model_type, submodel_type)
+
+    # Install the requested model.
+    model_installer.heuristic_import(model_path_id_or_url)
+
+    try:
+        return model_installer.mgr.get_model(model_name, base_model, model_type, submodel_type)
+    except ModelNotFoundException as e:
+        raise Exception(
+            "Failed to get model info after installing it. There could be a mismatch between the requested model and"
+            f" the installation id ('{model_path_id_or_url}'). Error: {e}"
+        )