This commit fixes bugs related to the on-the-fly conversion and loading of
legacy checkpoint models built on SD-2.0 base.
- When legacy checkpoints built on SD-2.0 models were converted
on-the-fly using --ckpt_convert, generation would crash with a
precision incompatibility error.
A long-standing issue with importing legacy checkpoints (both ckpt and
safetensors) is that the user has to identify the correct config file,
either by providing its path or by selecting which type of model the
checkpoint is (e.g. "v1 inpainting"). In addition, some users wish to
provide custom VAEs for use with the model. Currently this is done in
the WebUI by importing the model, editing it, and then typing in the
path to the VAE.
To improve the user experience, the model manager's
`heuristic_import()` method has been enhanced as follows:
1. When initially called, the caller can pass a config file path, in
which case it will be used.
2. If no config file provided, the method looks for a .yaml file in the
same directory as the model which bears the same basename. e.g.
```
my-new-model.safetensors
my-new-model.yaml
```
The yaml file is then used as the configuration file for
importation and conversion.
3. If no such file is found, then the method opens up the checkpoint
and probes it to determine whether it is V1, V1-inpaint or V2.
If it is a V1 format, then the appropriate v1-inference.yaml config
file is used. Unfortunately there are two V2 variants that cannot be
distinguished by introspection.
4. If the probe algorithm is unable to determine the model type, then its
last-ditch effort is to execute an optional callback function that can
be provided by the caller. This callback, named `config_file_callback`
receives the path to the legacy checkpoint and returns the path to the
config file to use. The CLI uses to put up a multiple choice prompt to
the user. The WebUI **could** use this to prompt the user to choose
from a radio-button selection.
5. If the config file cannot be determined, then the import is abandoned.
The user can attach a custom VAE to the imported and converted model
by copying the desired VAE into the same directory as the file to be
imported, and giving it the same basename. E.g.:
```
my-new-model.safetensors
my-new-model.vae.pt
```
For this to work, the VAE must end with ".vae.pt", ".vae.ckpt", or
".vae.safetensors". The indicated VAE will be converted into diffusers
format and stored with the converted models file, so the ".pt" file
can be deleted after conversion.
No facility is currently provided to swap a diffusers VAE at import
time, but this can be done after the fact using the WebUI and CLI's
model editing functions.
- This PR adds support for embedding files that contain a single key
"emb_params". The only example I know of this format is the
"EasyNegative" embedding on HuggingFace, but there are certainly
others.
- This PR also adds support for loading embedding files that have been
saved in safetensors format.
- It also cleans up the code so that the logic of probing for and
selecting the right format parser is clear.
- resolve conflicts with generate.py invocation
- remove unused symbols that pyflakes complains about
- add **untested** code for passing intermediate latent image to the
step callback in the format expected.
This PR fixes#2951 and restores the step_callback argument in the
refactored generate() method. Note that this issue states that
"something is still wrong because steps and step are zero." However,
I think this is confusion over the call signature of the callback, which
since the diffusers merge has been `callback(state:PipelineIntermediateState)`
This is the test script that I used to determine that `step` is being passed
correctly:
```
from pathlib import Path
from invokeai.backend import ModelManager, PipelineIntermediateState
from invokeai.backend.globals import global_config_dir
from invokeai.backend.generator import Txt2Img
def my_callback(state:PipelineIntermediateState, total_steps:int):
print(f'callback(step={state.step}/{total_steps})')
def main():
manager = ModelManager(Path(global_config_dir()) / "models.yaml")
model = manager.get_model('stable-diffusion-1.5')
print ('=== TXT2IMG TEST ===')
steps=30
output = next(Txt2Img(model).generate(prompt='banana sushi',
iterations=None,
steps=steps,
step_callback=lambda x: my_callback(x,steps)
)
)
print(f'image={output.image}, seed={output.seed}, steps={output.params.steps}')
if __name__=='__main__':
main()
```
This PR fixes#2951 and restores the step_callback argument in the
refactored generate() method. Note that this issue states that
"something is still wrong because steps and step are zero." However,
I think this is confusion over the call signature of the callback, which
since the diffusers merge has been `callback(state:PipelineIntermediateState)`
This is the test script that I used to determine that `step` is being passed
correctly:
```
from pathlib import Path
from invokeai.backend import ModelManager, PipelineIntermediateState
from invokeai.backend.globals import global_config_dir
from invokeai.backend.generator import Txt2Img
def my_callback(state:PipelineIntermediateState, total_steps:int):
print(f'callback(step={state.step}/{total_steps})')
def main():
manager = ModelManager(Path(global_config_dir()) / "models.yaml")
model = manager.get_model('stable-diffusion-1.5')
print ('=== TXT2IMG TEST ===')
steps=30
output = next(Txt2Img(model).generate(prompt='banana sushi',
iterations=None,
steps=steps,
step_callback=lambda x: my_callback(x,steps)
)
)
print(f'image={output.image}, seed={output.seed}, steps={output.params.steps}')
if __name__=='__main__':
main()
```
- This PR turns on pickle scanning before a legacy checkpoint file
is loaded from disk within the checkpoint_to_diffusers module.
- Also miscellaneous diagnostic message cleanup.
- When a legacy checkpoint model is loaded via --convert_ckpt and its
models.yaml stanza refers to a custom VAE path (using the 'vae:'
key), the custom VAE will be converted and used within the diffusers
model. Otherwise the VAE contained within the legacy model will be
used.
- Note that the heuristic_import() method, which imports arbitrary
legacy files on disk and URLs, will continue to default to the
the standard stabilityai/sd-vae-ft-mse VAE. This can be fixed after
the fact by editing the models.yaml stanza using the Web or CLI
UIs.
- Fixes issue #2917
- The value of png_compression was always 6, despite the value provided to the
--png_compression argument. This fixes the bug.
- It also fixes an inconsistency between the maximum range of png_compression
and the help text.
- Closes#2945
Prior to this commit, all models would be loaded with the extremely unsafe `torch.load` method, except those with the exact extension `.safetensors`. Even a change in casing (eg. `saFetensors`, `Safetensors`, etc) would cause the file to be loaded with torch.load instead of the much safer `safetensors.toch.load_file`.
If a malicious actor renamed an infected `.ckpt` to something like `.SafeTensors` or `.SAFETENSORS` an unsuspecting user would think they are loading a safe .safetensor, but would in fact be parsing an unsafe pickle file, and executing an attacker's payload. This commit fixes this vulnerability by reversing the loading-method decision logic to only use the unsafe `torch.load` when the file extension is exactly `.ckpt`.
Cause of the problem was inadvertent activation of the safety checker.
When conversion occurs on disk, the safety checker is disabled during loading.
However, when converting in RAM, the safety checker was not removed, resulting
in it activating even when user specified --no-nsfw_checker.
This PR fixes the problem by detecting when the caller has requested the InvokeAi
StableDiffusionGeneratorPipeline class to be returned and setting safety checker
to None. Do not do this with diffusers models destined for disk because then they
will be incompatible with the merge script!!
Closes#2836
Some schedulers report not only the noisy latents at the current timestep,
but also their estimate so far of what the de-noised latents will be.
It makes for a more legible preview than the noisy latents do.
* Fix img2img and inpainting code so a strength of 1 behaves the same as txt2img.
* Make generated images identical to their txt2img counterparts when strength is 1.
There are actually two Stable Diffusion v2 legacy checkpoint
configurations:
1) "epsilon" prediction type for Stable Diffusion v2 Base
2) "v-prediction" type for Stable Diffusion v2-768
This commit adds the configuration file needed for epsilon prediction
type models as well as the UI that prompts the user to select the
appropriate configuration file when the code can't do so
automatically.
This PR fixes the following scripts:
1) Scripts that can be executed within the repo's scripts directory.
Note that these are for development testing and are not intended
to be exposed to the user.
configure_invokeai.py - configuration
dream.py - the legacy CLI
images2prompt.py - legacy "dream prompt" retriever
invoke-new.py - new nodes-based CLI
invoke.py - the legacy CLI under another name
make_models_markdown_table.py - a utility used during the release/doc process
pypi_helper.py - another utility used during the release process
sd-metadata.py - retrieve JSON-formatted metadata from a PNG file
2) Scripts that are installed by pip install. They get placed into the venv's
PATH and are intended to be the official entry points:
invokeai-node-cli - new nodes-based CLI
invokeai-node-web - new nodes-based web server
invokeai - legacy CLI
invokeai-configure - install time configuration script
invokeai-merge - model merging script
invokeai-ti - textual inversion script
invokeai-model-install - model installer
invokeai-update - update script
invokeai-metadata" - retrieve JSON-formatted metadata from PNG files
This is the first phase of a big shifting of files and directories
in the source tree.
You will need to run `pip install -e .` before the code will work again!
Here's what's in the current commit:
1) Remove a lot of dead code that dealt with checkpoint and safetensor loading.
2) Entire ckpt_generator hierarchy is now gone!
3) ldm.invoke.generator.* => invokeai.generator.*
4) ldm.model.* => invokeai.model.*
5) ldm.invoke.model_manager => invokeai.model.model_manager
6) In addition, a number of frequently-accessed classes can be imported
from the invokeai.model and invokeai.generator modules:
from invokeai.generator import ( Generator, PipelineIntermediateState,
StableDiffusionGeneratorPipeline, infill_methods)
from invokeai.models import ( ModelManager, SDLegacyType
InvokeAIDiffuserComponent, AttentionMapSaver,
DDIMSampler, KSampler, PLMSSampler,
PostprocessingSettings )
This bug is related to the format in which we stored prompts for some time: an array of weighted subprompts.
This caused some strife when recalling a prompt if the prompt had colons in it, due to our recently introduced handling of negative prompts.
Currently there is no need to store a prompt as anything other than a string, so we revert to doing that.
Compatibility with structured prompts is maintained via helper hook.
- fix unused variables and f-strings found by pyflakes
- use global_converted_ckpts_dir() to find location of diffusers
- fixed bug in model_manager that was causing the description of converted
models to read "Optimized version of {model_name}'
Fixed a couple of bugs:
1. The original config file for the ckpt file is derived from the entry in
`models.yaml` rather than relying on the user to select. The implication
of this is that V2 ckpt models need to be assigned `v2-inference-v.yaml`
when they are first imported. Otherwise they won't convert right. Note
that currently V2 ckpts are imported with `v1-inference.yaml`, which
isn't right either.
2. Fixed a backslash in the output diffusers path, which was causing
load failures on Linux.
Remaining issues:
1. The radio buttons for selecting the model type are
nonfunctional. It feels to me like these should be moved into the
dialogue for importing ckpt/safetensors files, because this is
where the algorithm needs help from the user.
2. The output diffusers model is written into the same directory as
the input ckpt file. The CLI does it differently and stores the
diffusers model in `ROOTDIR/models/converted-ckpts`. We should
settle on one way or the other.
Converted the picker options to a Radio Group and also updated the backend to use the appropriate config if it is a v2 model that needs to be converted.
1) Downgrade numpy to avoid dependency conflict with numba
2) Move all non ldm/invoke files into `invokeai`. This includes assets, backend, frontend, and configs.
3) Fix up way that the backend finds the frontend and the generator finds the NSFW caution.png icon.
