search

huggingface / diffusers

🤗 Diffusers: State-of-the-art diffusion models for image and audio generation in PyTorch and FLAX.
https://huggingface.co/docs/diffusers
Apache License 2.0
25.97k stars 5.35k forks source link

Deteriorated Quality for DPM++ 2M Generations #7331

Open hipsterusername opened 7 months ago

hipsterusername commented 7 months ago

Describe the bug

In Invoke's 3.6.3 version, we updated to the latest Diffusers available at the time (.26) for the benefit of the scheduler fixes.

A user has provided documentation/testing demonstrating that these updates seems to have resulted in some worsened quality for certain schedulers (tested evidence is for DPM++ 2M)

Below --

Pre .26 32f59ecb-c909-4a96-9d6b-2523b933ce29 b7ae3ea8-95db-44af-a04e-210a08517419

.26 9d8389f7-4f7c-4af2-a416-29b860176524 89e840ea-0611-4250-ac07-26d57a1cf79f

Reproduction

Gen an SD1.5 image pre .26 and post .26 with DPM++ 2M

Logs

No response

System Info

Diffusers .27

Who can help?

@sayakpaul @yiyixuxu

yiyixuxu commented 7 months ago

@hipsterusername hey can you provide a reproducible script?

github-actions[bot] commented 6 months ago

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.

sayakpaul commented 4 months ago

Maybe @asomoza has some suggestions here?

asomoza commented 4 months ago

I would need more information, like the prompt, steps and cfg to do some tests.

Also this is kind of subjective, probably it was because the user was used to the results before the schedulers fix. I'm not that used to the results of SD 1.5 but for me the pre .26 have more artifacts and also a black spot which makes them more real but that kind of artifacts were the ones that were fixed.

Also the results post .26 look more plastic which for me is the normal kind of result with SD 1.5 and they don't have the artifacts from before so they look more clean but at the same time more fake.

AFAIK this change can be reverted with the scheduler config, just need to change the final_sigmas_type configuration and the user should get the results from before but can't check if I don't have generation info.

github-actions[bot] commented 3 weeks ago

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.