This repository accompanies the research paper "What Matters in Practical Learned Image Compression".
Authors: Kedar Tatwawadi, Parisa Rahimzadeh, Zhanghao Sun, Zhiqi Chen, Ziyun Yang, Sanjay Nair, Divija Hasteer, Oren Rippel
We introduce PICO (Perceptual Image Codec), the first learned codec that is both practical, and optimized directly for the human visual system. To derive it, we perform a comprehensive study of modeling choices for practical learned codecs, searching over millions of model configurations to jointly optimize over perceptual quality and on-device runtime.
Based on large-scale subjective user studies, PICO provides 2.3-3× bitrate savings against AV1, AV2, VVC, ECM and JPEG-AI, and 20-40% bitrate savings against the best learned codec alternatives. At the same time, on an iPhone 17, it encodes 12MP images as fast as 230ms, and decodes them in 150ms — faster than most top ML-based codecs run on a V100 GPU. Different from most learned codecs, PICO furthermore comes with cross-platform robustness guarantees.
The above figure shows comparisons of state-of-the-art traditional and learned codecs. Perceptual BD-rates are based on human ratings from the large-scale subjective study found in the paper. Speed benchmarks on iPhone 17 use identical compiler optimizations.
Visit our project page for an interactive tool that allows you to compare PICO to other codecs side-by-side.
We share here PICO reconstructions on non-PII images in the CLIC 2020 Test Set, across 8 different bitrates. These are the reconstructions directly from the subjective studies reported in the paper.
This software and accompanying data have been released under the following licenses:
- Code: Apple Sample Code License (ASCL)
- Data: CC-BY-NC-ND Deed
If you find our work useful, please cite:
@article{tatwawadi2026pico,
title={What Matters in Practical Learned Image Compression},
author={Tatwawadi, Kedar and Rahimzadeh, Parisa and Sun, Zhanghao and Chen, Zhiqi and Yang, Ziyun and Nair, Sanjay and Hasteer, Divija and Rippel, Oren},
journal={arXiv preprint arXiv:2605.05148},
year={2026}
}