Is it possible to migrate this to Stable Diffusion 3?

[Hzj199]

First of all, I would like to sincerely appreciate your impressive work on.
I have a question regarding the potential compatibility of this work. Would it be possible to migrate this approach to the Stable Diffusion 3 model? If so, are there any specific considerations or adjustments that would be necessary?
Looking forward to your guidance and insights. Thanks again for this outstanding contribution!

[ZiyiZhang27]

Thank you for your interest in our work! Yes, it is indeed possible to adapt our approach to Stable Diffusion 3. However, there are a few considerations and adjustments that would likely be necessary:

• Our approach is specifically designed to align few-step models efficiently. As such, it would be important to leverage a distilled few-step version of Stable Diffusion 3, such as Stable Diffusion 3.5 Large Turbo.

• Migrating to Stable Diffusion 3 may require upgrading our codebase to support newer versions of libraries like Accelerate, Diffusers, and Transformers.

If you'd like to work on this, feel free to reach out. I would be happy to help.

[Hzj199]

Thank you for sharing your insights! I am currently experimenting with this, but I’m uncertain how to implement a function similar to ddim_step_with_logprob in a flow-based model. I have made similar modifications based on diffusion-dpo and spo, but unfortunately, it hasn’t worked as expected.

Would you have any recommendations or insights on how to approach this?