How do you keep an image generator from drifting off course?

Autoregressive image and video generators learn from ground-truth sequences during training but must sample their own outputs at test time—like a student who practiced with answers but must work alone on the exam. This exposure bias causes drift and quality loss. Visual Prefix Guidance (VPG) fixes this at inference by contrasting predictions made with the generated sequence versus a corrupted one, then steering the model toward outputs that better support what it has already generated. No retraining needed.