The Journey, Not the Destination: How Data Guides Diffusion Models

Check out our blog post!

In our paper, we introduce a data attribution framework for diffusion models, together with an efficent method fo computing attribution scores. Given a generated image X and a diffusion model of interest, you can use our library to identify training examples which strongly guide the diffusion model towards generating X.

In particular, we provide attribution scores for each step of the diffusuion process:

Usage

Check out the examples. There, we:

• provide pre-computed attribution features so you can quickly score your generated images
• showcase how to compute the final scores using pre-computed features
• provide scripts to compute attribution features

Our code is based on the TRAK API.

Citation

If you use this code in your work, please cite using the following BibTeX entry:

@inproceedings{georgiev2023journey,
      title={The Journey, Not the Destination: How Data Guides Diffusion Models}, 
      author={Kristian Georgiev and Joshua Vendrow and Hadi Salman and Sung Min Park and Aleksander Madry},
      booktitle = {Arxiv preprint arXiv:2312.06205},
      year={2023},
}