Cool-chic 5.0 codec cuts encoding time 10x, beats H.266 by 11%

Overfitted image codecs take a contrarian approach to compression: instead of training one universal decoder, they learn a specialized decoder for each individual image at encode time. This trades heavy computational cost during encoding for strong compression ratios and extremely lightweight decoding. Cool-chic 5.0, the latest iteration from Orange's research team, tackles the main weakness of this approach—slow encoding—with a redesigned neural decoder architecture and an improved optimization process.

The new algorithm outperforms all previous overfitted codecs while requiring 10 times fewer encoding iterations, and delivers an 11% bitrate reduction against the state-of-the-art conventional codec H.266/VVC. Against modern learned autoencoders like MLIC++, Cool-chic 5.0 achieves competitive rate-distortion performance but with decoding complexity that is 250 times lower. That combination makes it particularly attractive for applications where decoding happens on resource-constrained devices, such as mobile phones, web browsers, or embedded systems. The code is open-source, allowing developers and researchers to integrate or build upon the new scheme for image delivery and storage use cases.