Cool-chic (pronounced /kul ʃik/ as in French 🥖🧀🍷) is a low-complexity neural image and video codec based on overfitting. It achieves state-of-the-art image compression performance, while requiring only 1000 multiplications per decoded pixel.
Cool-chic 5.0 focuses on image coding, implementing the contributions presented in the following paper: Cool-chic 5.0: Faster Encoding and Inter-Feature Entropy Modeling for Overfitted Image Compression, Ladune et al..
- 10x encoding speed-up vs. Cool-chic 4.2
- Better image compression: Rate decrease of -10% versus Cool-chic 4.2 for identical quality
- Low decoding complexity of 2000 MAC / pixel
Check-out the release history to see previous versions of Cool-chic.
See the Cool-chic setup documentation for additional details
# Clone the Cool-chic source code
git clone https://github.com/Orange-OpenSource/Cool-Chic.git && cd Cool-Chic
# Install create and activate virtual env
sudo apt install -y pip
python3 -m pip install virtualenv
python3 -m virtualenv venv && source venv/bin/activate
# Install required packages
pip install -r requirements.txt
# Sanity check
python -m test.sanity_checkSee the example page for more encoding examples. A comprehensive description of the encoder parameters are available for image and video compression.
# Encode an image using the default configuration
python cc_encode.py -i image.png -o ./bitstream.cool
# Video requires to successively encode multiples frames.
python samples/encode.py -i myTestVideo_1920x1080_24p_yuv420_8b.yuv -o bitstream.cool# Decoder outputs either PNG (image) or YUV (video) files
python cc_decode.py -i samples/bitstreams/kodim14.cool -o kodim14.png| BD-rate of Cool-chic 5.0.1 vs. [%] | ||||||
|---|---|---|---|---|---|---|
| Cheng | ELIC | MLIC++ | Cool-chic 4.2 | Cool-chic 5.0.0 | VVC (VTM 28.3) | |
| kodak (RGB) | -10.4 % | +0.1 % | +10.2% | -6.1 % | -1.2 % | -3.9 % |
| clic20-pro-valid (RGB) | -21.3 % | -9.5 % | +0.5 % | -9.9 % | -0.6 % | -11.6% |
| jvet A (YUV420) | / | / | / | / | -1.4 % | -4.0 % |
| jvet B (YUV420) | / | / | / | -13.7% | -1.9 % | +6.6 % |
If you use this project, please cite the Cool-chic 5.0 paper in your work.
@article{ladune2026-coolchic5-0-fasterencoding,
title={Cool-chic 5.0: Faster Encoding and Inter-Feature Entropy Modeling for Overfitted Image Compression},
author={Théo Ladune and Pierrick Philippe and Pierre Jaffuer and Théophile Blard and Sylvain Kervadec and Félix Henry and Gordon Clare},
year={2026},
eprint={2605.02726},
archivePrefix={arXiv},
primaryClass={eess.IV},
url={https://arxiv.org/abs/2605.02726},
}We rely on Robert Bamler's constriction package for range coding. Thanks to the developers for their work!
