How to skip unnecessary computation in image generation without losing quality?

Visual autoregressive models generate images one token at a time, but most tokens are redundant computation. Instead of guessing which ones to skip using shallow heuristics, this work measures each token's true impact on the final image by tracking changes in the model's internal state. On the Infinity-8B model, the method achieves 2.35× speedup while preserving generation quality, and requires no retraining.