MixCompress: Mixture of Experts for Variable Rate Learned Image Compression

2026-07-15

AI Summary

This work addresses the limitations of existing learned image compression methods that either require separate models for each bitrate or suffer from feature entanglement and performance degradation due to shared backbones in variable-bitrate approaches. To overcome these issues, the authors propose MixCompress, a novel framework that introduces sparse gated Mixture-of-Experts (MoE) into image compression for the first time to mitigate gradient conflicts. MixCompress further integrates dynamic depth expansion (Mixture-of-Depths, MoD) with Conditional Auxiliary Transform (CAT), enabling on-demand computational resource allocation and adaptive subband modulation. Through a hierarchical dynamic capacity scaling mechanism, MixCompress achieves inference efficiency while matching or surpassing the performance of single-bitrate specialized models, thereby establishing a new Pareto frontier in learned image compression.

Abstract

Learned image compression (LIC) is bottlenecked by the need to store independent models for each rate-distortion operating point. Existing variable bit-rate (VBR) methods aim to reduce this overhead via dense parameter modulation, but forcing a shared backbone to approximate divergent mappings causes severe feature entanglement. Specifically, low-rate smoothing gradients inherently conflict with the preservation of high-frequency textural details, leading to sub-optimal performance. To resolve this, we propose MixCompress, a unified VBR framework based on sparse structural specialization. While sparsely gated Mixture-of-Experts (MoE) routing successfully mitigates gradient conflict, it operates on a fixed computational budget. To address the increased representational demands of higher bit-rates we introduce a Mixture-of-Depths (MoD) extension to dynamically scale model capacity. Combined with Conditional Auxiliary Transforms (CAT) for dynamic sub-band energy modulation, our hierarchical framework effectively dynamically scales capacity. Extensive evaluations demonstrate that MixCompress not only matches individually optimized single-rate baselines but can even surpass them, establishing a new Pareto frontier for computationally efficient image coding.

Problem

• learned image compression
• variable bit-rate
• feature entanglement
• rate-distortion trade-off
• Mixture of Experts

Innovation

• Mixture-of-Experts
• Mixture-of-Depths
• Variable Bit-Rate Compression
• Learned Image Compression
• Conditional Auxiliary Transforms