Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling
The paper argues VAR image models lose local coherence because same-scale tokens are decoded in parallel.
Logit Refiner adds a small autoregressive module that samples those tokens sequentially while using frozen backbone features. The authors say it can attach to pretrained VAR checkpoints without retraining, adding about 10% parameters and under 5% of base training compute. In ImageNet 256x256 tests across 310M-to-2B-parameter backbones, the refiner improves generation quality and lets a 1.1B model beat one twice its size. They also report the same issue and fix carrying over to text-to-image VAR variants. ArXiv · AI/CL/LG's note
Logit Refiner adds a small autoregressive module that samples those tokens sequentially while using frozen backbone features. The authors say it can attach to pretrained VAR checkpoints without retraining, adding about 10% parameters and under 5% of base training compute. In ImageNet 256x256 tests across 310M-to-2B-parameter backbones, the refiner improves generation quality and lets a 1.1B model beat one twice its size. They also report the same issue and fix carrying over to text-to-image VAR variants. ArXiv · AI/CL/LG's note
score 5