Skip to content

Release IAM safety weights and demo on Hugging Face#1

Description

@NielsRogge

Hi @basim-azam 馃

I'm Niels and work as part of the open-source team at Hugging Face. I discovered your work on Arxiv and noticed that it has been accepted to ECCV 2026鈥攃ongratulations! 馃帀

The paper page on Hugging Face (https://huggingface.co/papers/2607.14945) allows users to discuss your paper and find associated artifacts. You can also claim the paper as yours, which will display it on your public profile on Hugging Face, and link your GitHub repository and project page.

I saw in your GitHub repository (https://github.com/basim-azam/iam) that the inference code and the learned safety direction weights (w_unsafe) are coming soon! Would you be interested in hosting these weights/models on https://huggingface.co/models?

Hosting on Hugging Face will provide much better visibility and discoverability for your work. We can add metadata tags to the model cards and directly link the repository to your paper page so that other researchers and developers can easily find and use it.

If you are interested, you can find a guide on uploading models to the Hub here. Since the weights represent a safety direction / modulation tensor, they can be easily fetched by users with a simple hf_hub_download one-liner.

In addition, since IAM is a training-free, inference-time control for FLUX models, it would be awesome to build an interactive demo of IAM on Hugging Face Spaces. We provide ZeroGPU which grants free, on-demand GPU-backed compute for Spaces, which would be perfect for showing the Light/Moderate/Strong safety trade-offs in real-time.

Let me know if you are interested or if you'd like any guidance on setting this up once you're ready to release!

Kind regards,

Niels

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions