MLICv2: Enhanced Multi-Reference Entropy Modeling for Learned Image Compression

A research team from Peking University and Peng Cheng Laboratory has unveiled MLICv2 and its enhanced variant MLICv2+, marking a significant leap in learned image compression (LIC). These models systematically address key limitations of previous AI-based codecs, particularly performance degradation at high bitrates and suboptimal entropy modeling. The announcement, detailed in a paper accepted to ACM TOMM, demonstrates that these successors to the MLIC series substantially outperform the current industry benchmark, Versatile Video Coding (VVC) Intra, setting a new state-of-the-art for compression efficiency.

The technical breakthroughs include a lightweight token mixing block inspired by MetaFormer architecture to boost transform capacity, and a novel hyperprior-guided global correlation prediction mechanism that captures global context from the very first slice of data. Combined with a channel reweighting module and exploration of Stochastic Gumbel Annealing for instance-specific optimization, the models achieve remarkable rate reductions of 16.54% to 24.35% across standard datasets like Kodak and CLIC Pro. This advancement signals a shift where AI-native codecs can consistently outperform traditional, hand-engineered standards, paving the way for more efficient image storage and transmission across the web and mobile networks.