RETHINKING CLASSIFIER-FREE GUIDANCE IN ON-POLICY DIFFUSION DISTILLATION

Modern diffusion models rely on Classifier-Free Guidance (CFG) to balance prompt adherence and image quality, typically by composing a positive and a negative velocity prediction. While On-Policy Distillation (OPD) is an effective way to train smaller student models, it is unclear how to best distill these two-branch CFG systems. This paper investigates whether matching the final composed velocity is sufficient or if it introduces hidden optimization failures.

The authors analyze the optimization dynamics of existing OPD methods that match the final CFG-composed velocity. They identify a phenomenon called Negative Branch Asymmetry (NBA), where the student model reduces the error in the positive branch by inadvertently increasing the error in the negative branch. This compensation is hidden at the training guidance scale but causes the model to drift or degrade when the guidance scale is adjusted at inference time. To fix this, the authors propose Positive-Direction Matching (PDM), which explicitly supervises the positive branch and the CFG conditional direction (the difference between positive and negative branches) as separate objectives.

• Branch Ambiguity: Naive CFG distillation is under-identified because the loss function only constrains the weighted sum of branch errors, allowing them to cancel each other out.
• NBA Emergence: When the teacher has access to privileged information in the negative branch (e.g., reference images), naive distillation fails to reduce both branch errors, leading to antagonistic dynamics where one branch improves at the expense of the other.
• Robustness: PDM effectively eliminates this compensation freedom. In dense-to-sparse video control tasks, PDM-trained models maintain high performance across varying inference guidance scales, whereas naive models exhibit severe sensitivity and visual distortion when the guidance scale deviates from the training value.