LoRA-powered image compression cuts parameters by 99% with variable rates
New method uses LoRA modules to achieve variable-rate compression with zero extra inference cost.
Deep Dive
A progressive learning approach for variable-rate deep image compression using Low-Rank Adaptation (LoRA) is proposed. The LoRA Rate-Adaptive Module (LoRAM) merges re-parameterized weights, adding no inference complexity. Compared to multi-model methods, it saves 99% in parameter storage, 90% in datasets, and 97% in training steps while achieving competitive performance.
Key Points
- Proposes LoRAM, a drop-in LoRA module for variable-rate deep image compression without extra inference complexity.
- Achieves 99% parameter storage reduction, 90% less training data, and 97% fewer training steps vs. multi-model methods.
- Competitive rate-distortion performance across diverse applications including web, medical, and connected vehicle networks.
Why It Matters
Enables highly efficient, adaptive image compression for bandwidth-limited systems like autonomous vehicles and remote diagnostics.