Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation
The paper identifies a branch-level failure mode in CFG-based on-policy distillation and proposes Positive-Direction Matching to fix it.
The authors argue that directly matching teacher and student guided velocities can hide compensating errors between the positive and negative CFG branches. They call the problematic case Negative Branch Asymmetry, where privileged information in the teacher’s negative branch can make positive-branch error fall while negative-branch error rises. Their proposed PDM objective separately constrains the positive prediction and the CFG conditional direction. They test it on dense-to-sparse video control, where it makes distillation less sensitive to inference guidance scales. ArXiv · AI/CL/LG's note
The authors argue that directly matching teacher and student guided velocities can hide compensating errors between the positive and negative CFG branches. They call the problematic case Negative Branch Asymmetry, where privileged information in the teacher’s negative branch can make positive-branch error fall while negative-branch error rises. Their proposed PDM objective separately constrains the positive prediction and the CFG conditional direction. They test it on dense-to-sparse video control, where it makes distillation less sensitive to inference guidance scales. ArXiv · AI/CL/LG's note
score 5