Wavefront parallelism speeds learned image compression 13x without retraining

Learned image compression has made great strides with autoregressive context models, but their sequential decoding is painfully slow—each pixel depends on previously decoded ones. Existing acceleration tricks like checkerboard context require architectural changes and full retraining, making them useless for pre-trained models. Now, Shimon Murai and colleagues from Waseda University propose a completely different approach: wavefront parallelization. Borrowed from video coding standards (e.g., H.264/HEVC), the method reorders inference into a staggered, diagonal wavefront. This allows multiple pixels to be processed in parallel while still respecting causal dependencies. The algorithm is applied at inference time only, no retraining needed. On Cheng et al.'s popular model, the team achieved a 13.4x speedup with zero loss in rate-distortion performance. They also show that by sacrificing some context fidelity, even faster decoding is possible. The paper has been accepted at MMSP 2026 and source code will be released.

This breakthrough matters because learned image compression has been held back by slow inference, especially for deployment on edge devices or real-time applications. Wavefront parallelization offers a plug-and-play speed boost for existing models, making state-of-the-art compression practical without the cost of retraining. The method is architecture-agnostic, so it can be applied to future autoregressive models as well. With the growing demand for efficient storage and transmission of visual data, this technique could accelerate adoption of learned compression in production systems—from cloud-based image servers to on-device photo apps. The trade-off option for even faster decoding adds flexibility for latency-sensitive use cases.