Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling

Visual Autoregressive Models (VAR) decode tokens within each scale in parallel, an efficiency choice that acts as a mean-field approximation and discards intra-scale spatial dependencies. Logit Refiner introduces a lightweight autoregressive plug-in that restores sequential sampling over frozen backbone features, requiring roughly 10% parameter overhead and under 5% base training compute. Across ImageNet benchmarks from 310M to 2B parameters, the method allows a 1.1B model to surpass an unrefined 2B baseline, proving that repairing decoding rules can rival brute-force scale.