HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Mon, 22 Dec 2025 22:44:45 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"6949c9dd-8489" expires: Sun, 28 Dec 2025 08:15:47 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: 200A:3946E9:758258:83C1D1:6950E4DB accept-ranges: bytes age: 0 date: Sun, 28 Dec 2025 08:05:48 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210082-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1766909148.844178,VS0,VE221 vary: Accept-Encoding x-fastly-request-id: d9287fc6238fcb5ba9082c30a683fc2f0ae7571d content-length: 7673

Overview

Can we use VLMs as judges to improve human-aligned image compression? Yes! In VLIC (Vision Language Models for Image Compression), we present a diffusion-based image compression system designed to be post-trained with binary VLM judgments. VLIC leverages existing techniques for diffusion model post-training with preferences, rather than distilling the VLM judgments into a separate perceptual loss network. Please consult our paper for more details, and check out the visualizations on this page!

We observe that VLIC produces high-quality reconstructions, particularly for human-relevant details such as text and faces. We conduct large-scale user studies and quantitative evaluations. Please consult the paper for the details. Thank you!

Acknowledgments

We thank Ben Poole, David Minnen, and Dina Bashkirova for helpful discussions.

BibTeX

        @article{sargent2025vlic,
            title   = {VLIC: Vision-Language Models As Perceptual Judges for Human-Aligned Image Compression},
            author  = {Sargent, Kyle and Gao, Ruiqi and Henzler, Philipp and Herrmann, Charles and Holynski, Aleksander and Li, Fei-Fei and Wu, Jiajun and Zhang, Jason},
            journal = {arXiv preprint arXiv:XXXX.XXXXX},
            year    = {2025}
        }