EF-LIC eliminates entropy coding, boosting image compression 5x faster decode

Traditional learned image compression relies on entropy coding to convert latent representations into a compact bitstream, but this step is sequential and becomes a major latency bottleneck. To overcome this, Hao Cao and colleagues from multiple institutions propose EF-LIC (Entropy-Coding Free Learned Image Compression), a multi-rate framework accepted at ICML 2026. Their approach introduces two key innovations: unconstrained vector quantization (UVQ) that pushes index distribution toward the maximum-entropy bound to minimize statistical redundancy, and a context-conditioned autoregressive transform that reparameterizes latents to reduce inter-dependency. Theoretical analysis proves EF-LIC can remove correlation redundancy as effectively as typical LIC with entropy coding.

On the Kodak dataset, EF-LIC achieves up to 67.86% bitrate reduction over the MS-ILLM baseline when measured with LPIPS, a perceptual metric. Ablation studies confirm it matches the compression performance of its entropy-coding-based variant while delivering over 3x faster encoding and 5x faster decoding. This breakthrough could dramatically reduce image compression latency for real-time applications such as video streaming, cloud gaming, and edge AI, where every millisecond matters. The work is particularly relevant for mobile and embedded devices that require low-power, high-speed compression without sacrificing quality.