Thank you for the new release

[nitinmukesh]

Any plans to integrate in diffusers just like v1
huggingface/diffusers#10148

With diffusers it becomes available for low VRAM users using
pipe.enable_sequential_cpu_offload()

and quantization like nf4.

I have v1 working on as low as 6-8 GB VRAM, slow but it works.

Thank you

[Luciennnnnnn]

Yes, we do plan to integrate OmniGen2 into diffusers, and it may be ready in the coming weeks.

[nitinmukesh]

Thank you.

[nitinmukesh]

Change the dress to blue, flowers petal to red and earrings to green.

[Luciennnnnnn]

Good try

[nitinmukesh]

Was testing prompt adherence and it worked 👍.
Kudos to the team and thank you for sharing with community.

[Luciennnnnnn]

@nitinmukesh Hi, cpu offload is supported now, you can try it directly in our repo. See this

[able2608]

Thanks for implementing sequential and model offloading. I recently played around the code a bit (also to implement offloading) and stumbled upon group_offloading provided by the diffusers library which claims to provide a middle ground between loading the whole model and layer offloading (by loading a group of layers and allow asynchronous parameter loading and execution). From their documentation it seems that they had provided a function to wrap any pytorch model with that functionality. I'm curious about whether such strategy had been tested by your team or if it is not feasible to use such functionality. Are there any technical barriers to using it from your testing?

[Luciennnnnnn]

Hi @able2608, thanks for the suggestion. We are aware of group_offloading, but we haven't explored it yet as its implementation appears to be quite complex than other cpu offloading strategies.

If you'd like to take a shot at implementing it, feel free to open a PR and we'll be glad to review it!

[able2608]

@Luciennnnnnn I have created a pull request regarding group offloading (and some other general fixes). Perhaps you can check it out and see whether it fits or not.

[nitinmukesh]

@nitinmukesh Hi, cpu offload is supported now, you can try it directly in our repo. See this

Thank you. :)

[ChunChenLin]

Yes, we do plan to integrate OmniGen2 into diffusers, and it may be ready in the coming weeks.

Hi, is OmniGen2 integrated into diffusers yet? (Since I didn't find it available at diffusers)
Thanks!