Opening the Black Box of Classifier-Free Guidance via Information Bottleneck
The paper investigates when to activate Classifier-Free Guidance (CFG) in diffusion models by treating it as a time-dependent information bottleneck. It shows that conditional information's effectiveness isn't constant but peaks after a critical stage governed by a Fisher-information balance condition. This leads to a training-free, dynamic CFG strategy that activates guidance based on monitoring score statistics.