Multi-scale Image Representation Compression

The researchers proposed the MIRC multi-scale image representation compression model, which uniformly quantizes all encoding components such as latent variables, synthetic networks, and entropy models, and performs end-to-end compression under a single rate-distortion objective. MIRC introduces multi-scale representations with cross-stage parameter sharing to optimize coding efficiency. On the CLIC2020 professional validation set, this model achieved a 10.5% reduction in BD-rate compared to VVC (VTM 22.0). Additionally, MIRC offers decoding configurations ranging from 1.2 to 2.9 kMAC per pixel to meet different deployment requirements.