arXiv 2605.090712026-09-21cs.CV

概率流蒸馏：用于高保真3D生成的精确Wasserstein梯度流

Probability-Flow Distillation: Distribution Matching in Parameter Space

Department of Physics（物理系）；Indian Institute of Technology Madras（印度理工学院马德拉斯分校）；Department of Electrical Engineering（电气工程系）

Rohith Ramanan, A. N. Rajagopalan

备注：: This version corrects an error in v1 that overlooked the implicit dependence of $q_0$ on the flow map and the confusion between practical Jacobian zeroing and its theoretical treatment under stop-gradient. Thus, v1 PFD is not an exact Wasserstein gradient descent. We strengthen the analysis, add new results, introduce a stronger PFD variant, and revise the title and abstract

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文提出概率流蒸馏，通过精确Wasserstein梯度流解决SDI的模式崩溃问题，实现高保真3D生成。

摘要：

Score Distillation Sampling (SDS)及其变体已被广泛用于文本到3D生成，通过蒸馏2D图像扩散先验。然而，标准SDS目标易导致严重模式崩溃，产生过度平滑和饱和的结果。尽管最近的进展如通过反向的Score Distillation via Inversion (SDI)减轻了这些伪影，但最终仍无法准确捕捉完整的目标分布。本文表明，限制SDI采样能力的瓶颈来自其依赖后验均值估计，这在数学上等价于确定性反向DDIM轨迹的单步欧拉近似。为了解决这个问题，我们提出了一种自然启发的扩展，称为概率流蒸馏（PFD）。我们证明PFD恰好对应于Wasserstein梯度流，从而诱导出原理上的分布匹配动态。最后，我们展示PFD能够生成具有精细、高保真的3D资产，并在质量上优于现有方法。

英文摘要：

Score distillation methods use pretrained diffusion models as priors for optimizing parameters through differentiable forward models, most notably in text-to-3D generation. Yet the distribution they induce over those parameters is not well understood. Observing that existing distillation methods reduce to one of three: Score Distillation Sampling (SDS), Score Distillation via Inversion (SDI), and Variational Score Distillation (VSD), we extend the particle variational inference view of VSD to the other two. We show that SDS collapses onto the modes of the target, while SDI converges to a contracted version of it, and explain why SDI needs a negative classifier-free guidance scale. Next, we observe that the DDIM posterior mean equals a single Euler step of the probability-flow ODE (PF-ODE). Replacing this step in SDI with a full reverse solve makes the target a fixed point, but it requires solving two concatenated PF-ODEs. Dropping a Jacobian from the resulting gradient gives Probability-Flow Distillation (PFD), which requires solving only the forward PF-ODE. Experiments on synthetic targets, the CelebA dataset, and text-to-3D generation support our analysis and demonstrate the practical effectiveness of PFD.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。