Awesome
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:
<div style="width:65%; margin: 0 auto;"> <table style="width:100%;"> <tr> <td><img src="assets/mscoco.gif" alt="MSCOCO Attributions GIF"></td> <td><img src="assets/cifar.gif" alt="CIFAR10 Attributions GIF"></td> </tr> </table> </div>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},
}