AI summaryⓘ
The authors explain that autoregressive image and video generators usually learn by predicting the next part based on correct previous parts during training, but during actual use they have to rely on their own past guesses, which can cause errors over time. They introduce a method called Visual Prefix Guidance (VPG) that helps the model make better predictions by comparing outputs from the current generated sequence with outputs from a slightly messed-up version and then adjusting predictions to support the sequence more strongly. This method works during generation without needing to retrain the model and improves the quality of generated images and videos in several tested systems. Essentially, the authors found a new way to make AI-generated pictures and videos more accurate by smartly guiding its guesses step-by-step.
autoregressive generationexposure biasprefix driftteacher forcinglogitsimage generationvideo generationFID (Fréchet Inception Distance)inference-time guidancetext-to-image generation
Authors
Xinyao Liao, Qiyuan He, Yicong Li, Jiayin Zhu, Xiaoye Qu, Wei Wei, Angela Yao
Abstract
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.