Rethinking CFG in OPD: branch-aware distillation fixes Negative Branch Asymmetry

Bingnan Li · hf · 2026-07-28

This paper rethinks classifier-free guidance in on-policy diffusion distillation (OPD).

It shows that directly matching guided predictions is under-identified at the branch level: positive and negative branch errors can cancel each other out in the composed CFG output.
In a shared-negative-conditioning setting, naive guided matching can still work because both branch errors shrink together.
But when the teacher’s negative branch contains privileged information unavailable to the student, the dynamics become antagonistic: improving the positive branch can worsen the negative one.
The authors name this failure mode Negative Branch Asymmetry (NBA).
To fix it, they propose Positive–Direction Matching (PDM), which constrains the positive prediction and the CFG conditional direction separately.
On dense-to-sparse video control, PDM is reported to be more robust to inference guidance scale changes than naive guided matching.

More from Multimodal