Search papers, labs, and topics across Lattice.
The paper introduces Visual Prefix Guidance (VPG), a novel training-free inference-time technique to mitigate exposure bias and prefix drift in autoregressive image and video generation. VPG enhances next-step prediction by contrasting model outputs under the generated prefix against outputs under a corrupted prefix, extrapolating logits to favor candidates that reinforce the generated prefix's posterior support. Experiments across VAR, Infinity, and InfinityStar demonstrate that VPG improves generation quality, reducing FID on VAR by 0.36 on average and improving benchmark performance on both image and video generation tasks.
Autoregressive image and video generators can get a surprising boost in quality *without* retraining, just by steering them towards outputs that better support their own past creations.
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.
