HuggingFace Daily Papers（社区热门论文）· HuggingFace Daily Papers（社区热门论文）·· 2026-08-04精选AI 评分76
Any-OPD：面向流匹配模型的异构同策略蒸馏框架
AI 导读
Any-OPD 提出首个支持任意异构流匹配生成器对的同策略蒸馏框架，仅通过冻结的 DINOv2 表示空间桥接教师与学生模型，无需共享 VAE、架构或噪声调度。将 FLUX.1-dev 蒸馏至 SD3.5-Medium 后，学生模型 PickScore 与 HPSv3 均显著提升，以教师五分之一的参数量达到接近其性能，而直接潜在回归训练完全失败。
推荐理由
将异构流匹配蒸馏转化为在独立视觉空间求解固定点，配合噪声级对齐跨过VAE与架构差异，2.5B学生模型在偏好指标上逼近12B教师。
正文
Abstract: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.
| Subjects: | Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV) |
| Cite as: | arXiv:2608.03316 [cs.LG] |
| (or arXiv:2608.03316v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2608.03316 arXiv-issued DOI via DataCite |
Submission history
From: Siming Fu [view email]
[v1]
Tue, 4 Aug 2026 08:23:57 UTC (1,243 KB)
来源：HuggingFace Daily Papers（社区热门论文） · arxiv.org