Efficient Learned Image Compression without Entropy Coding

A new image compression framework eliminates the sequential entropy coding step that typically bottlenecks speed, while matching the performance of conventional learned methods, according to a preprint posted to arXiv on 22 May 2026 [1]. The approach, called Entropy-Coding Free Learned Image Compression (EF-LIC), replaces entropy coding with two mechanisms designed to remove statistical and correlation redundancy directly from the latent representation [1]. The authors introduce unconstrained vector quantization and prove mathematically that the resulting index distribution approaches the maximum-entropy bound, which they state yields minimal statistical redundancy [2]. A context-conditioned autoregressive transform then reparameterizes the latents to reduce inter-dependency among them [2]. Entropy coding is a standard component in typical learned image compression pipelines, where it converts latent representations into compact bitstreams [2]. Because the process is inherently sequential, it often becomes the dominant source of coding latency [1]. By removing this step, EF-LIC produces a multi-rate framework that the researchers report achieves over 3× faster encoding and 5× faster decoding compared with an entropy-coding-based variant of the same model [2]. On the Kodak dataset, EF-LIC delivered up to a 67.86% bitrate reduction relative to MS-ILLM when evaluated with the LPIPS perceptual metric [1]. Ablation studies further indicated that the entropy-coding-free design matched the compression performance of its entropy-coding counterpart while providing the latency gains [2]. The work sits within the broader field of deep learning, which employs multilayered neural networks for tasks including representation learning and computer vision [4]. Modern learned compression systems often rely on transformer-style architectures, a class of models that became prominent after the 2017 “Attention Is All You Need” paper demonstrated gains in parallelizability over earlier recurrent models [3]. EF-LIC’s autoregressive transform draws on related principles to condition each latent on its context, though the paper does not claim to use transformer blocks directly [2]. The preprint has not yet been peer-reviewed. The authors posted it to the Electrical Engineering and Systems Science section of arXiv under the Image and Video Processing subcategory [1].