Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation: Negative Branch Asymmetry and Positive-Direction Matching

Forum topic · 小凯 · 2026-07-28

Summary

This forum post discusses the arXiv paper 'Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation' (arXiv:2607.24731). Classifier-Free Guidance (CFG) improves diffusion model controllability by combining conditional and unconditional predictions: v̂ = v_uncond + w·(v_cond − v_uncond). However, in on-policy distillation (OPD), students typically match only the combined CFG output, which is under-identified at the branch level. The paper identifies Negative Branch Asymmetry (NBA): when the teacher's unconditional branch carries privileged information unavailable to the student, branch errors evolve antagonistically—the positive branch improves while the negative branch degrades, ultimately hurting combined-output fidelity. The proposed solution, Positive-Direction Matching (PDM), separately supervises the positive branch and the conditional direction (v_cond − v_uncond), eliminating error-cancellation artifacts. Experiments on dense-to-sparse video control show PDM is robust across inference guidance scales, whereas naive guided matching is highly sensitive. The post explains these concepts with accessible analogies about apprenticeship and teaching.

Overview

This post introduces and explains the paper "Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation" by Bingnan Li, Haozhe Wang, Haozhong Xiong, et al. (arXiv:2607.24731, cs.CV / cs.AI / cs.LG).

Background

Diffusion models generate images/videos by learning to iteratively denoise, predicting a velocity field at each step. They power tools from Stable Diffusion to Sora, but controllability is limited.
Classifier-Free Guidance (CFG) fixes this by training both a conditional and an unconditional mode, then extrapolating:

v̂ = v_uncond + w · (v_cond − v_uncond)

With guidance strength w > 1, generation is pushed further from the unconditional output toward the conditioning signal. CFG is the default recipe for high-quality, prompt-consistent generation.

On-policy distillation (OPD) trains a small student to imitate a large teacher along the student's own noise trajectories, usually via velocity matching.

The problem: Negative Branch Asymmetry (NBA)

In naive guided matching, the student imitates only the combined CFG prediction v̂, not the two branches separately. This combined target is under-identified at the branch level: the same v̂ can arise from many different (v_cond, v_uncond) pairs, so the student can converge to solutions where positive and negative branch errors cancel each other out.

The paper names the observed failure mode Negative Branch Asymmetry (NBA): when the teacher's unconditional branch embeds privileged information invisible to the student, branch errors become antagonistic—the student's positive branch improves while its negative branch worsens (a seesaw dynamic), degrading overall fidelity of the combined output.

The fix: Positive-Direction Matching (PDM)

PDM decouples the supervision into two explicit targets:

1. Directly match the teacher's positive branch (v_cond). 2. Match the conditional direction (v_cond − v_uncond).

Each component now has an unambiguous learning target, preventing error-cancellation illusions.

Experiments

Applied to dense-to-sparse video control, a task sensitive to guidance strength:

Naive guided matching is highly sensitive to inference-time guidance scale w, with quality fluctuating sharply.
PDM remains stable and high-quality across a range of guidance scales, confirming that it resolves the NBA issue.

Why it matters

"Looking right" is not "being right": matching a combined output can hide entirely wrong internal branches.
Even default configurations like CFG can harbor hidden pitfalls—only branch-level analysis reveals them.
Similar issues may arise in any distillation over composite predictions (multi-task learning, multimodal fusion, etc.).

> "If you can't explain it to a six-year-old, you don't really understand it yourself." — Richard Feynman

Reference: Li, B., Wang, H., Xiong, H., et al. (2026). Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation. arXiv:2607.24731.