MixCompress: Mixture of Experts for Variable Rate Learned Image Compression

Learned image compression (LIC) typically requires training and storing separate models for every target bit-rate, which is computationally expensive and inefficient for deployment. Existing variable bit-rate (VBR) methods attempt to solve this by using a single shared backbone with dense parameter modulation. However, these methods often suffer from feature entanglement, where the model struggles to simultaneously optimize for low-rate smoothing and high-rate detail preservation, leading to sub-optimal performance compared to single-rate models.

The authors propose MixCompress, a unified VBR framework that replaces dense modulation with sparse structural specialization. The core components include:

• Mixture-of-Experts (MoE): Sparsely gated expert networks that isolate gradient updates for different rate points, mitigating the conflict between low-rate and high-rate optimization objectives.
• Mixture-of-Depths (MoD): An extension that dynamically scales model capacity by routing higher bit-rate inputs to deeper, more complex expert networks.
• Conditional Auxiliary Transforms (CAT): A rate-aware wavelet-based shortcut that modulates sub-band energy compaction based on the target bit-rate, allowing for adaptive feature refinement without altering the entropy model.

MixCompress effectively decouples the optimization trajectories of competing rate points, as evidenced by a shift in gradient cosine similarity toward orthogonality. By training a single model for the same duration as a single-rate baseline, the authors demonstrate that MixCompress can match or surpass the performance of state-of-the-art single-rate models like LALIC and LIC-TCM. This establishes a new Pareto frontier for image coding, providing high representational density and superior rate-distortion performance without increasing inference costs.