Any-OPD: Heterogeneous On-Policy Distillation for Flow-Matching Models via Representation-Space Bridging

📄 arXiv: 2608.03316v1 📥 PDF

作者: Siming Fu, Zheming Fu, Ruizhe He, Hualiang Wang, Jie Huang, Xiaoxiao Ma, Mingchen Zhong, Weihu Huang, Xiaoxuan He, Haojun Xu

分类: cs.LG, cs.CV

发布日期: 2026-08-04


💡 一句话要点

提出Any-OPD以解决异构模型间的在线蒸馏问题

🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture)

关键词: 在线蒸馏 潜在流匹配 模型压缩 计算机视觉 生成模型 黑箱采样 模型无关表示

📋 核心要点

  1. 现有的在线蒸馏方法假设教师和学生模型具有相同的潜在空间和架构,但在实际应用中常常不成立,导致性能下降。
  2. Any-OPD框架通过将教师视为黑箱采样器,利用模型无关的视觉表示进行样本比较,解决了模型间的潜在不匹配问题。
  3. 在将12B的FLUX.1-dev蒸馏到2.5B的SD3.5-Medium中时,Any-OPD将学生模型的PickScore从0.846提升至0.884,HPSv3从9.12提升至10.97。

📝 摘要(中文)

在线蒸馏中,教师模型需要纠正学生模型生成的样本,然而当教师和学生来自不同模型家族时,传统方法无法有效应用。本文提出Any-OPD框架,首次实现任意配对的潜在流匹配生成器之间的在线蒸馏。该方法将教师视为黑箱采样器,通过在冻结的、模型无关的视觉表示中比较独立解码的输出,避免了对潜在变量、特征或架构的假设。实验表明,Any-OPD显著提升了学生模型的性能,展现出强大的应用潜力。

🔬 方法详解

问题定义:论文要解决的问题是如何在教师和学生模型来自不同家族时进行有效的在线蒸馏。现有方法在潜在空间和架构不匹配时,无法提供有效的目标,导致训练失败或性能下降。

核心思路:论文提出的Any-OPD框架将教师模型视为黑箱,通过在一个冻结的、模型无关的视觉表示中比较学生和教师的输出,避免了对潜在变量和特征的假设,从而实现了有效的蒸馏。

技术框架:Any-OPD的整体架构包括两个主要模块:首先是教师样本的重编码阶段,使用学生模型的VAE对教师样本进行重编码;其次是通过匹配连续噪声水平而非时间步索引来恢复轨迹对应关系。

关键创新:最重要的技术创新在于将教师模型视为黑箱采样器,并通过模型无关的视觉表示进行比较,这与传统方法依赖于潜在变量和架构的假设有本质区别。

关键设计:在设计中,Any-OPD采用了重编码阶段以确保样本质量的度量,而不是领域不匹配,此外,使用连续噪声水平匹配来替代传统的时间步索引,确保了蒸馏过程的有效性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果显示,Any-OPD在将12B的FLUX.1-dev模型蒸馏到2.5B的SD3.5-Medium模型时,学生模型的PickScore从0.846提升至0.884,HPSv3从9.12提升至10.97,展现出与教师模型相媲美的性能,且模型规模仅为教师的五分之一。

🎯 应用场景

该研究的潜在应用领域包括计算机视觉、生成模型和强化学习等。通过实现不同模型间的有效蒸馏,Any-OPD可以帮助在资源受限的环境中部署高效的生成模型,具有重要的实际价值和广泛的应用前景。

📄 摘要(原文)

On-policy distillation, in which a teacher corrects samples that the student itself generates, presupposes that the two models speak the same language: identical VAE latents, matching architectures, and a common timestep grid. We ask what happens when none of this holds, as when the strongest teacher available and the student one wishes to deploy come from different model families, and find that the standard recipes have no answer: teacher latents cannot serve as targets in a foreign coordinate system, per-pixel losses against a teacher that stochastically re-draws local detail degenerate into blur or divergence, and timestep indices lose their meaning across mismatched schedules. We present Any-OPD, to our knowledge the first framework for on-policy distillation between arbitrary pairs of latent flow-matching generators. Any-OPD treats the teacher purely as a black-box sampler and connects the two models at exactly one point: a frozen, model-agnostic vision representation in which their independently decoded outputs are compared, sidestepping every assumption about latents, features, or architecture. Trajectory correspondence is recovered by matching continuous noise levels instead of step indices, and a brief anchoring phase, in which teacher samples are re-encoded through the student's own VAE, ensures the on-policy gradient measures sample quality rather than domain mismatch. Distilling the 12B FLUX.1-dev into the 2.5B SD3.5-Medium, Any-OPD lifts the student's PickScore from 0.846 to 0.884 and HPSv3 from 9.12 to 10.97, rivaling the teacher at a fifth of its size, where direct latent regression fails to train at all.