[论文] Posterior Matching Distillation
论文概要
研究领域: CV
作者: Hidir Yesiltepe, Julius Berner, Arash Vahdat, Pinar Yanardag
发布时间: 2026-11-19
arXiv: 2611.26629
中文摘要
扩散蒸馏通过将预训练教师的迭代采样过程压缩为少量网络评估来加速图像和视频生成。基于分布的方法通过将学生输出分布与教师匹配来实现这一目标,这种匹配通常由反向KL目标驱动,其模式寻求偏差以牺牲样本多样性为代价偏好主导模式。我们提出后验匹配蒸馏(PMD),这是分布蒸馏家族的新成员,将覆盖模式的后验采样与任意步数生成相结合。在零总体损失下,PMD在每个训练噪声水平恢复教师干净图像的完整后验。其随机学生支持在训练时间步网格内任意步数的采样预算,无需额外训练。为学习这些后验,PMD将学生样本的指数重加权平均值(以固定噪声状态为条件)与对应教师后验均值匹配。在高斯噪声下,指数倾斜后验均值仅通过在偏移输入上的一次教师评估即可计算。我们证明,在开邻域上匹配倾斜均值可唯一确定完整后验。因此,PMD通过一族均值约束监督整个条件分布,无需任何辅助学生分数网络。在SDXL和FLUX.2 4B上的实验表明,PMD在两个模型上均取得4步方法中的最佳FID,同时在提示对齐和感知质量指标上保持竞争性或领先的性能。
原文摘要
Diffusion distillation accelerates image and video generation by compressing a pretrained teacher's iterative sampling process into a few network evaluations. Distribution-based methods achieve this by matching the student's output distribution to the teacher's. This matching is often driven by reverse-KL objectives, whose mode-seeking bias favors dominant teacher modes at the expense of diversity across samples. We introduce Posterior Matching Distillation (PMD), a new member of the distribution-based distillation family that combines mode-covering posterior sampling with any-step generation. At zero population loss, PMD recovers the teacher's full posterior of clean images at each trained noise level. Its stochastic student supports any-step sampling budgets within the trained timestep g...
自动采集于 2026-11-21
#论文 #arXiv #CV #小凯