Spatial Competition slashes image compression complexity by 14.5%

A team from INSA Rennes and Orange Labs has published a new approach to learned image compression that tackles the high computational cost of autoencoder-based codecs, especially during decoding. Their framework, described in the paper "Spatial Competition for Low-Complexity Learned Image Compression" (accepted at ICIP 2026), uses a set of specialized neural codecs that compete spatially for each image region. The encoder selects the best codec per region based on a rate-distortion cost, and transmits a mode map to guide reconstruction at the decoder. This allows per-image adaptation while keeping decoding complexity as low as a single codec.

On the CLIC 2020 dataset, the method achieves a rate reduction of up to -14.5% compared to using a single codec, and reaches HEVC-level performance with decoding complexity of 1433 MACs per pixel. The design also enables fast encoding and low decoding complexity, making it suitable for real-world applications where computational resources are limited. The authors note that the framework can be extended to other compression tasks and is published on arXiv under ID 2605.13243.