Support for UNet3DConditionModel

[neurowelt]

Thank you so much for this repository! As I am currently working with both image and video pipelines, I added the UNet 3D regsiter methods, so you could use FreeU in Zeroscope/Modelscope models.

Changelog:

• Add UpBlock3D and CrossAttnUpBlock3D register methods
• Add scale argument required for newer diffusers versions (tested on v.0.21.2)
• Cosmetics, added method descriptions, improved logging and cleaned up imports

[charlesmccarthy]

Amazing please join us on the t2v and zeroscope discord https://discord.gg/hQEjvgRQkY

[YisuiTT]

Thank you so much for this repository! As I am currently working with both image and video pipelines, I added the UNet 3D regsiter methods, so you could use FreeU in Zeroscope/Modelscope models.

Changelog:

• Add UpBlock3D and CrossAttnUpBlock3D register methods
• Add scale argument required for newer diffusers versions (tested on v.0.21.2)
• Cosmetics, added method descriptions, improved logging and cleaned up imports

Hello, I use this code in animatediff, but there is an error：TypeError: register_free_upblock3d..up_forward..forward() got an unexpected keyword argument 'encoder_hidden_states'

[neurowelt]

@YisuiTT Could you provide a bit more context - which AnimateDiff are you using and how do you register the FreeU blocks?

[YisuiTT]

Thanks for answer. I use this AnimateDiff：https://github.com/tumurzakov/AnimateDiff, and add FreeU like this:

the whole error：

[neurowelt]

Thanks for the traceback @YisuiTT , I'll try to reproduce it and will let know when I figure it out 🤗

[neurowelt]

@YisuiTT Sorry for not checking this out for such a long time, it was a bit hectic for me. I think you should check out the newest diffusers update, it contains a lot of things including AnimateDiff and FreeU.

You can check out the detailed changelog here.

[YisuiTT]

🆗, thanks for telling me.