Image compression: does AI break every record with MLICv2?

Imagine sending an ultra-high-res photo in a flash, without losing quality. It sounds like magic, but it is the latest milestone of AI applied to image compression. A new study introduces MLICv2 and its beefed-up version MLICv2+, two models that outperform today's most advanced video standard, VVC Intra.

Until now, learned compression systems struggled at high quality: fine details were lost, and information handling wasn't optimal at the start of the process. Researchers fixed this with three clever moves. First, they added a lightweight "token mixing" block inspired by the MetaFormer architecture: a sort of efficient mixer that blends global info without heavy compute. Second, they taught the model to predict global context from the very first data slice using hyperprior guidance, adding a module that dynamically weights the most informative channels. Third, they explored per-image optimization via Stochastic Gumbel Annealing (SGA), adapting compression to specific content.

The numbers speak volumes: on famous datasets like Kodak, Tecnick, and CLIC, MLICv2+ cuts the required bitrate by 20-24% compared to the VTM-17.0 reference. That means smaller files, smoother streaming, and lighter archives, without sacrificing a single pixel. Published on arXiv in April 2025, this research sets a new state of the art: AI no longer mimics traditional codecs, it surpasses them by writing its own rules.

The MLICv2 architecture introduces targeted innovations addressing historical bottlenecks in Learned Image Compression (LIC). The MetaFormer-inspired Token Mixing Block (MTMB) replaces classic self-attention or deep convolutions with an efficient token mixing operation (often MLP or global pooling based), solving high-bitrate degradation caused by limited transform capacity in prior models (e.g., ELIC, MLICv1).

On the entropy modeling front, Hyperprior-Guided Global Correlation Prediction (HGGP) tackles the "first slice" problem: in multi-slice autoregressive architectures (checkerboard or channel-wise), the first slice lacks lateral context. HGGP uses the decoded hyperprior to predict a global correlation map, injecting otherwise missing context. The Channel Reweighting Module (CRM) then applies a Squeeze-and-Excitation style recalibration on Gaussian parameters (mean/scale) per channel, suppressing noise on less informative channels.

Enhanced Positional Embedding improves spatial awareness in lightweight transformers, while Guided Selective Compression enables non-uniform bit allocation guided by importance maps. Finally, Stochastic Gumbel Annealing (SGA) replaces classic additive uniform noise during training with annealed Gumbel-Softmax sampling, enabling instance-specific optimization at test time (test-time adaptation) that refines quantization parameters for the specific image. BD-Rate gains (-16% to -24% vs VTM-17.0 Intra) confirm superiority over VVC Intra even in high-fidelity regimes (PSNR/MS-SSIM).