VPG: Visual Prefix Guidance for Autoregressive Image and Video Generation
For practitioners using autoregressive generative models, VPG offers a lightweight, training-free fix to improve generation quality without retraining.
Autoregressive image and video generators suffer from exposure bias and prefix drift during inference. The authors propose Visual Prefix Guidance (VPG), a training-free inference-time method that improves next-step prediction by contrasting outputs under generated and corrupted prefixes, reducing FID by 0.36 on VAR and improving benchmark performance across multiple tasks.
Autoregressive image and video generators are trained with teacher-forced histories but must sample from their own generated prefixes at inference time, making them vulnerable to exposure bias and prefix drift. Existing remedies either modify training or apply sampling-time guidance aimed primarily at external semantic conditions, such as class labels or text prompts, rather than testing whether a next-step prediction provides strong posterior support for the generated prefix itself. We propose Visual Prefix Guidance (VPG), a training-free inference-time guidance method for autoregressive image and video generation. VPG improves next-step prediction by contrasting the model's output under the generated prefix with its output under a corrupted prefix, then extrapolating logits toward candidates that strengthen the posterior support of the generated prefix. Across class-conditional image generation with VAR, text-to-image generation with Infinity, and text-to-video generation with InfinityStar, VPG improves generation quality without retraining the base model, reducing FID on VAR by 0.36 on average and improving benchmark performance on both image and video generation.