Paper 2607.24731
Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation
- Published
- Jul 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 8 stars
01 In brief
Summary
This paper analyzes on-policy distillation (OPD) for diffusion models under classifier-free guidance (CFG).
The authors show that the naive objective of matching CFG-composed velocities is under-identified at the branch level, allowing positive- and negative-branch errors to cancel.
They identify a failure mode, Negative Branch Asymmetry (NBA), which occurs when the teacher's negative branch contains privileged information unavailable to the student, causing antagonistic branch-error dynamics and excess sensitivity to inference guidance scale.
To address this, they propose Positive–Direction Matching (PDM), which separately constrains the positive prediction and the CFG conditional direction, and compare it with Independent Branch Matching (IBM).
Experiments in text-rendering and reference-conditioned image distillation confirm NBA's emergence under privileged conditioning.
In dense-to-sparse video control with Wan-VACE, PDM improves control fidelity across pose, depth, and scribble modalities and remains stable across guidance scales, unlike naive matching.
Ablations show λ=1 and K=8 as effective hyperparameters.
The authors conclude that branch-aware supervision prevents cross-branch error compensation and improves robustness, though the theoretical advantage of PDM over IBM remains open.
02 From the paper
Abstract
On-policy distillation (OPD) adapts diffusion models by querying a teacher along trajectories generated by the current student, but how it should behave under classifier-free guidance (CFG), a default component of modern diffusion systems, remains poorly understood. Existing OPD methods naturally extend velocity matching to the CFG-composed prediction, directly matching teacher and student guided velocities. We show that this objective is under-identified at the branch level: positive- and negative-branch errors can compensate in the guided prediction. Through two contrasting cases, we find that naive matching remains effective under shared negative conditioning, where both branch errors decrease jointly. When the model's native CFG schema retains privileged information in the teacher's negative branch that is unavailable to the student, however, this joint reduction breaks down and the composed objective induces antagonistic branch-error dynamics, reducing the positive-branch error while increasing the negative-branch error. We term this failure mode Negative Branch Asymmetry (NBA). To address NBA, we introduce Positive--Direction Matching (PDM), a branch-aware OPD objective that separately constrains the positive prediction and the CFG conditional direction. We apply PDM to dense-to-sparse video control, where naive guided matching is highly sensitive to inference guidance scales, while branch-aware supervision enables more robust and effective knowledge transfer.