---
title: "Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.24731v1"
permalink: "https://feed.astralor.com/items/itm_2502596a980056125090073e"
published: "2026-07-27T17:57:02.000Z"
score: 79
channel: "ai"
tags: ["Classifier-Free Guidance","On-policy Distillation","Diffusion Model Distillation","Positive-Direction Matching","Negative Branch Asymmetry"]
---

# Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation

## AI commentary

> This paper hits a subtle but critical issue in diffusion model distillation—the negative branch asymmetry under classifier-free guidance—with clear theoretical analysis and solid experiments. A must-read for engineers and researchers working on diffusion model fine-tuning or knowledge distillation.

## At a glance

On-policy distillation (OPD) adapts diffusion models by querying a teacher, but the role of classifier-free guidance (CFG) remains poorly understood. Existing methods directly match guided velocities, yet this objective suffers from under-identification at the branch level: positive and negative branch errors can compensate. When the teacher's negative branch contains privileged information unavailable to the student, naive matching induces Negative Branch Asymmetry (NBA). The authors propose Positive-Direction Matching (PDM), a branch-aware objective that separately constrains the positive prediction and CFG direction. Experiments on dense-to-sparse video control demonstrate more robust knowledge transfer.

## Key points

- Existing on-policy distillation methods directly match guided velocities, but suffer from under-identification: positive and negative branch errors can compensate.
- Identifies Negative Branch Asymmetry (NBA) failure mode when teacher's negative branch contains privileged information not available to the student.
- Proposes Positive-Direction Matching (PDM), a branch-aware objective that separately constrains the positive prediction and CFG conditional direction.
- Validates on dense-to-sparse video control, achieving more robust knowledge transfer across inference guidance scales.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_2502596a980056125090073e