Efficient Learned Image Compression without Entropy Coding

Entropy coding is widely used in typical learned image compression (LIC) that converts latents into a compact bitstream. However, entropy coding is typically sequential and becomes the coding latency bottleneck. To overcome it, we present **E**ntropy-Coding **F**ree **L**earned **I**mage **C**ompression (EF-LIC), a multi-rate framework that generates compact representation by removing statistical and correlation redundancy with low coding latency. First, we introduce unconstrained vector quantization and prove that its index distribution approaches the maximum-entropy bound, yielding minimal statistical redundancy. Second, we propose a context-conditioned autoregressive transform that directly reparameterizes the latents to reduce inter-dependency. Theoretical analysis shows that EF-LIC can remove correlation redundancy as effectively as typical LIC with entropy coding, leading to comparable compression performance. Experiments show EF-LIC achieves up to 67.86\% bitrate reduction over MS-ILLM on Kodak with LPIPS. Ablation studies further show EF-LIC matches the compression performance of its entropy-coding based variant while achieving over $3\times$ faster encoding and $5\times$ faster decoding.

Learned image compression aims to store images with fewer bits while keeping them visually faithful. Most existing methods rely on entropy coding, a final coding step that makes the stored representation compact, but this step is usually sequential and often becomes the latency bottleneck. This paper asks whether learned image compression can remain efficient without entropy coding. We propose EF-LIC, an entropy-coding-free image compression framework that directly learns compact image representations. The key idea is to reduce two types of redundancy before storage: statistical redundancy, where some codes are used much more often than others, and correlation redundancy, where parts of the representation repeat information already contained in other parts. EF-LIC uses learned discrete codes and a context-conditioned transform to remove these redundancies while keeping the whole codec friendly to parallel GPU execution. Experiments show that EF-LIC achieves strong visual compression performance and substantially lower encoding and decoding latency than several existing methods. This makes it a promising direction for fast, deployable image compression on devices, servers, and real-time visual systems.