Possible Bug in DDIMScheduler when prediction_type = 'sample'

[JianxinMa]

Describe the bug

https://github.com/huggingface/diffusers/blob/v0.9.0/src/diffusers/schedulers/scheduling_ddim.py#L303
https://github.com/huggingface/diffusers/blob/v0.9.0/src/diffusers/schedulers/scheduling_ddim.py#L278

It seems that the ``model_output'' at L303 should be the predicted epsilon. However, it is actually the predicted original sample when prediction_type == "sample".

A workaround is to use ``use_clipped_model_output=True''. And the results are much worse if use_clipped_model_output=False, even when clip_sample=False.

Reproduction

No response

Logs

No response

System Info

• diffusers version: 0.9.0
• Platform: macOS-12.6-x86_64-i386-64bit
• Python version: 3.9.7
• PyTorch version (GPU?): 1.10.0 (False)
• Huggingface_hub version: 0.10.1
• Transformers version: 4.23.1
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

[patrickvonplaten]

Hey @JianxinMa,

Could you maybe copy-paste a reproducible code snippet below? :-)

[PeterL1n]

Also encountered that bug. @patrickvonplaten I opened a PR to fix it. Please approve. #1788

[patrickvonplaten]

Commented on the PR - thanks for opening it!

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.