[Feature] Add support for HiDream-I1

[Aunali321]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/mit-han-lab/nunchaku/discussions/new/choose. Otherwise, it will be closed.
• 2. I will do my best to describe the issue in English.

Motivation

HiDream-I1 is the SOTA open source image gen model but it requires around 80GB of VRAM.

There is NF4 version of it here that runs on 15GB of VRAM. But it would be nice to have original-like quality that SVDQuant usually provides.

Related resources

Original: https://github.com/HiDream-ai/HiDream-I1

Community Improved: https://github.com/hykilpikonna/HiDream-I1-nf4

[vlrevolution]

I'm also supporting this endevour, please could we have support for HiDream-l1?

[nitinmukesh]

HiDream-I1
+1

[vlrevolution]

Just look at this prompt adherence:

[CHNtentes]

It would be better if this model can somehow fit in my 12GB vram. Currently it seems not possible...

[nitinmukesh]

Someone posted comparison with FLUX
https://www.reddit.com/r/StableDiffusion/comments/1jw6z42/some_hidreamdev_nf4_comfy_vs_fluxdev_comparisons/

Also diffusers integration in progress
huggingface/diffusers#11231

Please consider this implementation.

Best part is
HiDream-ai/HiDream-I1 is licensed under the
MIT License

[vedicjyotishuniverse]

Please add support of this model, thank you.

[J-Cott]

It will be amazing if they do this, Hi-Dream is good, but it is currently sooo slow compared to Flux Nunchaku!

[Foul-Tarnished]

In the meantime, use the GGUF from City96 GGUF Loader on ComfyUI

Should be usable on 12GB

(Q6_K barely fit in my 16GB vram gpu)

For speed, add SageAttention, Triton, and torch.compile, all on nightly pytorch, for a big speed boost (3.6s/it to 2.2s/it)
Don't forget to add --use-sage-attention