Thanks to the inspiring progress in Visual Autoregressive (VAR) models!
We introduce SAR (Self-Autoregressive Refinement), a lightweight post-training framework that fixes the train–test mismatch in scale-wise autoregressive image generation.
Instead of RL-based alignment, SAR shows that self-rollout alone can serve as a robust and stable post-training strategy.
With Stagger-Scale Rollout and a Contrastive Student-Forcing Loss, SAR enables effective error correction across scales and consistently improves image quality with minimal overhead (up to 5.2% FID↓).
Arxiv: https://arxiv.org/abs/2512.06421
Project Page: https://gengzezhou.github.io/SAR/
Github: https://gengzezhou.github.io/SAR/
Thanks to the inspiring progress in Visual Autoregressive (VAR) models!
We introduce SAR (Self-Autoregressive Refinement), a lightweight post-training framework that fixes the train–test mismatch in scale-wise autoregressive image generation.
Instead of RL-based alignment, SAR shows that self-rollout alone can serve as a robust and stable post-training strategy.
With Stagger-Scale Rollout and a Contrastive Student-Forcing Loss, SAR enables effective error correction across scales and consistently improves image quality with minimal overhead (up to 5.2% FID↓).
Arxiv: https://arxiv.org/abs/2512.06421
Project Page: https://gengzezhou.github.io/SAR/
Github: https://gengzezhou.github.io/SAR/