custom CFG?

[jtoy]

is it possible to Set CFG to zero after 40% of steps?
I see in this article he does it: https://www.baseten.co/blog/sdxl-inference-in-under-2-seconds-the-ultimate-guide-to-stable-diffusion-optimiza/

He has sample code, but I dont know how to incorporate it.
I couldn't find how to do this in the documentation. he is doing it with diffusors library.

[a-r-r-o-w]

I believe the code mentioned in the article is to be incorporated here.

Essentially, to perform CFG, you need to duplicate the latents and prompt embeddings. In order to turn it off after a certain number of inference steps, like the article mentions, you remove the duplication of latents by setting do_classifier_free_guidance=False and halving the prompt embeds (which was initially duplicated in encode_prompt()).

[patrickvonplaten]

We're working on supporting this! cc @yiyixuxu

See: #4778

[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.