Logit Refiner Restores Spatial Dependencies in Visual Autoregressive Models

The Logit Refiner is a lightweight autoregressive module that improves Visual Autoregressive (VAR) model outputs by modeling intra-scale dependencies. It adds ~10% parameters and <5% training compute to any pretrained VAR checkpoint to fix local incoherence in models ranging from 310M to 2B parameters.