Add HF integration to MAR

[jadechoghari]

Hello @LTH14 and team, Congratulations on this amazing work—MAR looks incredible! 🎉

You should consider adding MAR to the Hugging Face Hub 🤗. Doing so could increase the model's visibility and make it easier for everyone to use. It can easily be made compatible with the Hugging Face Transformers library.

I've already started the process (almost done 🚀), and other folks from HF will be available to help. You can check the progress here: https://huggingface.co/jadechoghari/mar

Soon, it will be easily usable with just one line of code. And will adda gradio demo as well!

Best regards, Jade

[LTH14]

Hi Jade,

Thank you for your interest and support! We’re happy to see MAR being integrated into the Hugging Face library. Please feel free to let us know if there’s anything we can do to help.

[jadechoghari]

superb, I'll let you know if any issue is encountered.

Btw thanks so much for providing a notebook, it will help a lot!

[jadechoghari]

@LTH14 , everything is done 🎊: https://huggingface.co/jadechoghari/mar

Please confirm if anything needs to be changed before it is officially added to Hugging Face (ReadMe, Demo, etc...). A Gradio demo will be added soon!

[LTH14]

Thank you very much! Everything looks perfect. Just two comments:

1. you could add cfg_schedule as an option in the usage example. Two possible options are "constant" and "linear", and the default one to generate good images can be "constant".
2. You could consider using "mar_huge" as default and use the images in the colab to replace the current one, so that the user can exactly reproduce the images

[jadechoghari]

done! for any other updates if easier for you, you can just open a PR in my repo

[LTH14]

Great! Also let me know when the Gradio demo is added, and I can add it to this repo also

[jadechoghari]

Of course!

[jadechoghari]

@LTH14 Gradio demo is done! 🤗 Check it out here: https://huggingface.co/spaces/jadechoghari/mar (running on A100s). Let u-s know if any changes are needed... Otherwise still waiting for MIT's HF org access approval. Also, will open a PR to add the HF model cards + Gradio demo to this repo if you want

Best!!

[LTH14]

Thanks a lot! It looks awesome! One comment is that, if I change the number of images generated, the 50.8s progress bar seems unchanged -- is it possible to use the number of iterations instead of the progress bar?

I have already added the huggingface and Gradio demo to the Readme. Again, thanks a lot for supporting it!

[LTH14]

Also -- what is "MIT's HF org access approval"?

[jadechoghari]

@LTH14 I believe Gradio averages the timestep after each use, which is why it's displayed on the progress bar. We could update the demo's README and mention the timesteps time, like: "64 steps = ~50 seconds." What do you think?

"Also -- what is "MIT's HF org access approval"?" -> we have an unofficial MIT org on HF (i thought it was official) : https://huggingface.co/MIT. It'd be great if your lab has any official org linked to MIT on HF, we could add it their & announce it!

[LTH14]

Then I think we can just leave it there and don't need to add it to readme. After all, small batch generation is much slower per sample than large batch because the small diffusion MLP cannot fully use the GPU utility.

That one seems unofficial -- please don't link to that one. We don't have an official org yet -- once we have it we can link mar to it

[jadechoghari]

alright - it's actually super easy to setup an org! finally It would be great to have this demo app.py locally available on the repository and linked in the readme. For example, like for Stable-fast-3D here: https://github.com/Stability-AI/stable-fast-3d?tab=readme-ov-file#local-gradio-app

I'll open a PR for it so you see what I mean!

edit: forgot to cc @LTH14 :)

[jadechoghari]

i think all done here just this PR left before closing the issue @LTH14 🚀 :)