huggingface
always discard next-to-last sigma for diffusers
Is your feature request related to a problem? Please describe. There's a feature from A1111 that I want to use in diffusers
Describe the solution you'd like. I'd like it implemented or instructions on how to implement it
Describe alternatives you've considered. figuring it out myself
Additional context. Add any other context or screenshots about the feature request here.
Cc: @yiyixuxu.
Can you give some results that benefit from this feature? That will help us to prioritise if this is worth the consideration.
@sayakpaul I have clear examples for you: https://cdn.discordapp.com/attachments/1115540317366657107/1184462530756743250/00054-1971414394.png?ex=658c0fb1&is=65799ab1&hm=e559596aac360775d665a00577add42147fdbc9a4a9b6999a7f562a15d05ad5c&
https://cdn.discordapp.com/attachments/1115540317366657107/1184462531465588906/00055-1971414394.png?ex=658c0fb1&is=65799ab1&hm=3ea2deaf5a2dff9c93febbd928f000de37655dfd44d8625ca7a5eeaa42e494f4&
It impacts the shadows. Look at the shoulders and face extremeties
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.
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
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.
It's easy to add, but I'm not convinced with the benefit. would love to see some results comparison
@yiyixuxu here's a result comparison:
https://cdn.discordapp.com/attachments/1115540317366657107/1184462530756743250/00054-1971414394.png?ex=658c0fb1&is=65799ab1&hm=e559596aac360775d665a00577add42147fdbc9a4a9b6999a7f562a15d05ad5c&
https://cdn.discordapp.com/attachments/1115540317366657107/1184462531465588906/00055-1971414394.png?ex=658c0fb1&is=65799ab1&hm=3ea2deaf5a2dff9c93febbd928f000de37655dfd44d8625ca7a5eeaa42e494f4&
the change is small but impactful
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.
cc @Beinsezii here let me know what you think about this, would it be helpful to support this?
The image links aren't valid anymore so I'm not sure what exactly it's doing.
Assuming it's something like
+ sigmas[-1] = 0 # nullify sigma while keeping vector length for indexing
sigmas = np.cat(sigmas[-1])
You essentially oversample the image at the end by progressing through the timesteps faster which results in a smoother image. IMO that's the wrong way to do it though, as it's much easier and more controllable to just adjust the magnitude of the input noise which is already possible in diffusers.
Magnitude 0.99, 1, 1.02 using my diffusers cli
If that's not it, you can probably still patch it in manually by performing an in-place adjustment of pipe.scheduler.sigmas before calling. As long as the tensor length doesn't shrink it shouldn't explode.
It would be nice to have some images attached directly to github so they aren't subject to Discord's whims
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.