Note: The target branch is `feat/ip-adapter`, not `main`. After a
cursory review here, I'll merge for an in-depth review as part of
https://github.com/invoke-ai/InvokeAI/pull/4429.
## Description
This branch adds model management support for IP-Adapter models. There
are a few notable/unusual aspects to how it is implemented:
- We have defined a model format that works better with our model
manager than the 'official' IP-Adapter repo, and will be hosting the
IP-Adapter models ourselves (See `invokeai/backend/ip_adapter/README.md`
for a description of the expected model formats.)
- The CLIP Vision models and IP-Adapter models are handled independently
in the model manager. The IP-Adapter model info has a reference to the
CLIP model that it is intended to be run with.
- The `BaseModelType.Any` field was added for CLIP Vision models, as
they don't have a clear 1-to-1 association with a particular base model.
## QA Instructions, Screenshots, Recordings
Install the following models via the InvokeAI UI:
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)
## What type of PR is this? (check all applicable)
- [x] Refactor
- [ ] Feature
- [ ] Bug Fix
- [ ] Optimization
- [ ] Documentation Update
- [ ] Community Node Submission
## Have you discussed this change with the InvokeAI team?
- [x] Yes
- [ ] No, because:
## Description
**NOTE!!!** This PR is against `feat/ip-adapter`, not `main`. I created
a PR because I made some pretty significant changes that I thought might
spark discussion.
I don't think it makes sense to do a full in-depth review here. If
possible, let's try to agree on the high-level approach and then merge
this and do an in-depth review on the original PR.
High-level changes:
- Split `IPAdapterField` from the `ControlField` and make them separate
inputs on the `DenoiseLatentsInvocation`
- Create context manager that handles patching/un-patching the UNet with
IP-Adapter attention blocks (`IPAdapter.apply_ip_adapter_attention()`)
- Pass IP-Adapter conditioning via `cross_attention_kwargs` rather than
concatenating it to the text embedding. This helps avoid breaking other
features (like long prompts).
- Remove unused blocks of the IP-Adapter implementation and do some
general tidying.
Out of scope:
- I haven't looked at model management yet. I'd like to get this merged
into `feat/ip-adapter` and then look at model management separately.
