We propose Adaptive Classifier-Free Guidance (A-CFG), which dynamically re-masks low-confidence tokens for a more targeted unconditional input in iterative models.