[论文] PDMD: Projected Distribution Matching Distillation for Video Diffusion...
研究领域: CV 作者: Zimo Wang, Junkun Yuan, Angtian Wang, Haotian Yang, Canyu Zhang, Siyuan Yuan, Xingchang Huang, Bo Liu, Yizhi Wang, Yiding Yang, Chongyang Ma, Gord…
论文概要
研究领域: CV 作者: Zimo Wang, Junkun Yuan, Angtian Wang, Haotian Yang, Canyu Zhang, Siyuan Yuan, Xingchang Huang, Bo Liu, Yizhi Wang, Yiding Yang, Chongyang Ma, Gordon Guocheng Qian 发布时间: 2026-09-28 arXiv: 2609.35768
中文摘要
现代视频扩散模型需要在长时空 token 序列上进行数十次去噪评估。分布匹配蒸馏(DMD)将函数评估次数(NFE)减少到仅几次。然而,DMD 样本在训练过程中会退化,表现出渐进式过饱和和伪影。我们将这种不稳定性追溯到评论器(critic)误差,它们进入连续的学生更新并随时间累积。我们提出投影分布匹配蒸馏(PDMD)来过滤评论器误差:将 DMD 更新中与学生-评论器端点残差平行的分量投影去除。在固定的含噪查询点,我们证明该残差是评论器端点误差的无偏估计。在高维假设下,该投影去除了恒定比例的评论器误差,同时仅丢弃趋于零的理想 DMD 信号。实证上,该投影稳定了训练并在 DMD 退化和产生不自然纹理的地方改善了样本质量。PDMD 只需对 DMD 做一行代码修改,无需额外损失、网络、数据、模型传递或多阶段训练。使用 Wan2.1,PDMD 在 4 NFE 下达到 83.73 的 VBench 总分,超过同等训练的 DMD 1.03 分。在 MiniMax-H3 联合视频-音频生成上,PDMD 达到 83.17 的 VideoGen-Eval 视觉总分,比最强蒸馏基线高 0.41 分,并在全部六个音频指标上取得最佳。用户研究在视觉质量、运动和音频质量方面均偏好 PDMD。
原文摘要
Modern video diffusion models require tens of denoising evaluations over long spatiotemporal token sequences. Distribution Matching Distillation (DMD) reduces the number of function evaluations (NFE) to just a few. However, DMD samples can degrade during training, exhibiting progressive oversaturation and artifacts. We trace this instability to critic errors, which enter successive student updates and accumulate over time. We introduce Projected Distribution Matching Distillation (PDMD) to filter critic errors. PDMD projects out the component of the DMD update parallel to the student-critic endpoint residual. At a fixed noisy query, we prove that this residual is an unbiased estimate of the critic's endpoint error. Under high-dimensional assumptions, this projection removes a constant frac...
*自动采集于 2026-09-30*
#论文 #arXiv #CV #小凯