huggingface / huggingface/diffusers

AnimateDiff SparseCtrl RGB does not work as expected

Open
#9,508 9 comments 0 reactions 0 assignees View on GitHub
advanced bug contributions-welcome help wanted stale
Dominant language
Python
Stars
34.5k
Forks
7.3k
Avg merge
3d 3h
Merged PRs (30d)
91

Description

Relevant comments are [this](https://github.com/huggingface/diffusers/pull/8897#issuecomment-2255416318) and [this](https://github.com/huggingface/diffusers/pull/8897#issuecomment-2255478105).

AnimateDiff SparseCtrl RGB does not work similar to other implementations and cannot replicate their outputs. This makes me believe that there is something incorrect with our SparseControlNet or MotionAdapter implementation.

When comparing the results of the [original](https://github.com/guoyww/AnimateDiff)/[Comfy](https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved) implementation to Diffusers implementation, one can notice that if an image is used with an unrelated prompt, the Diffusers implementation ignores the image and just follows the prompt whereas the other implementations try to incorporate both.

Since the original and Comfy implementations produce this behaviour consistently, this seems more like a problem with Diffusers implementation. However, I've not been able to spot differences in implementation just by comparing the code visually. I also tried matching outputs layerwise and it seemed to be alright (although I didn't investigate this as deeply as I should have due to other priorities).

If someone from the community actively following/using the AnimateDiff implementations can help determine the cause of this bug, it would be really awesome and helpful.

Contributor guide

Open the contributing guide

Research direction

Start with the linked comments on PR 8897 and reproduce the AnimateDiff SparseCtrl RGB comparison using the original and Comfy implementations as references. Investigate the Diffusers SparseControlNet and MotionAdapter behavior with an unrelated prompt, and compare the conditioning results. Done means identifying and correcting the cause so the image influences the output rather than being ignored.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
computer-vision, machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.