Reproduce “MixCompress: Mixture of Experts for Variable Rate Learned Image Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Calvin-Khang Ta, Praneet Singh, Tong Shao, Peng Yin
Dolby Laboratories developed MixCompress, a unified variable bit-rate learned image compression framework that integrates Mixture-of-Experts/Depths and Conditional Auxiliary Transforms to mitigate gradient interference. This approach achieves BD-rate reductions (PSNR) of up to -28.88% on Tecnick compared to VTM-23.1, often surpassing individually optimized single-rate models while significantly reducing model deployment overhead.
Key insights
- Gradient interference, caused by conflicting optimization objectives for different bit-rates in dense VBR models, is a primary cause of performance degradation.
- Sparse structural specialization using Mixture-of-Experts (MoE) or Mixture-of-Depths (MoD) can effectively mitigate this interference by routing features through specialized pathways.
- Dynamically scaling representational capacity via Mixture-of-Depths (MoD) is more effective than solely parameter-based specialization for adapting to diverse rate-distortion demands.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.