Variable-Rate Deep Image Compression based on Low-Rank Adaptation by Progressive Learning
Researchers have developed a novel approach to variable-rate deep image compression using Low-Rank Adaptation (LoRA). This method introduces a LoRA Rate-Adaptive Module (LoRAM) that allows a single model to achieve different compression rates without increasing computational complexity during inference. Experiments show this technique offers competitive performance while significantly reducing parameter storage, dataset requirements, and training steps compared to multi-model approaches. AI