State-of-the-art LIC models process entire feature maps as network inputs, making peak memory a hard bottleneck for high-resolution content. Block-based processing mitigates this but triggers blocking artifacts.

We mathematically derive the minimum overlap required to reproduce Full-image LIC results exactly, modeling how overlap propagates layer-by-layer through a CNN using a recursive formula. Several implementation techniques are applied on top of this theoretical foundation.

The result: artifact-free reconstruction with zero BD-rate loss compared to Full-image inference, while significantly reducing peak memory and peak MACs across 2K and 4K resolutions.

We design lightweight backbone blocks deployable in two separate contexts.

SCM (for NNVC / standard codec tools): A Spatial-then-Channel Mixing block applied to NN-based super-resolution and in-loop filtering. The NN-based super-resolution tool has been adopted into the NNVC codec software, and the in-loop filtering tool has been advanced to the Exploration Experiment (EE) stage within the JVET standardization process — a formal milestone toward standard inclusion.

SC-Gate (for LIC): Eliminates all global modeling (Self-attention, Mamba, Bi-RWKV) and uses a single depth-wise convolution for spatial mixing combined with an element-wise gating mechanism. At 28.8% of LALIC’s per-block complexity, it achieves –15.84% BD-rate vs. VTM 9.1, with gains validated on high-resolution datasets (Tecnick 1K, CLIC up to 2K).

MSE-optimized codecs produce over-smoothed, perceptually unrealistic reconstructions at low bitrates. The key insight is that a compressed image should (1) lie on the manifold of real images (Realism) while (2) remaining consistent with the original (Fidelity).

We mathematically formulate the transform and quantization characteristics of AI-based codecs and derive explicit constraints for diffusion model sampling. This constrained diffusion decoding keeps reconstructions on the real-image manifold without sacrificing semantic consistency — outperforming existing generative codecs in fidelity while matching their realism.