MixCompress beats image compression trade-offs with sparse expert routing
New MoE-MoD framework matches single-rate performance at all bitrates.
Learned image compression (LIC) faces a fundamental bottleneck: each rate-distortion operating point requires a separate model, leading to heavy storage overhead. Existing variable bit-rate (VBR) methods try to share a single backbone through dense parameter modulation, but this causes severe feature entanglement. Low-rate smoothing gradients conflict with high-frequency texture preservation, resulting in sub-optimal performance.
To solve this, Calvin-Khang Ta, Praneet Singh, Tong Shao, and Peng Yin propose MixCompress, accepted at ECCV 2026. Their key innovation is using sparsely gated Mixture-of-Experts (MoE) routing to isolate gradient conflicts, combined with a Mixture-of-Depths (MoD) extension that dynamically scales model capacity for higher bitrates. Conditional Auxiliary Transforms (CAT) further enable dynamic sub-band energy modulation. Evaluations show MixCompress not only matches individually optimized single-rate baselines but can even surpass them, delivering a new Pareto frontier for computationally efficient image coding.
- Sparse MoE routing mitigates gradient conflict between low-rate smoothing and high-rate texture preservation.
- Mixture-of-Depths (MoD) dynamically scales model capacity to meet representational demands of higher bitrates.
- Conditional Auxiliary Transforms (CAT) enable dynamic sub-band energy modulation for fine-grained rate control.
Why It Matters
Reduces model storage and compute overhead while achieving state-of-the-art compression quality across all bitrates.