Skip to content

Conversation

@xkvma
Copy link

@xkvma xkvma commented Dec 17, 2025

Returns NABLA attention to 10sec Kandinsky5 models.

Speeds up generation by ~2.7x time, without significant drop in quality

@kabachuha
Copy link

Is it possible to add NABLA to all the other models as well training-free? Would be a great addition

@xkvma
Copy link
Author

xkvma commented Dec 18, 2025

Is it possible to add NABLA to all the other models as well training-free? Would be a great addition

We really want this too, but it won't give performance gain as is.

Hope to do further work in future to let it work training-free.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants