用最优传输改进单步生成，3D点云更快更准
OT-MeanFlow3D: Bridging Optimal Transport and Meanflow for Efficient 3D Point Cloud Generation
- 结合最优传输与均值流，实现单步生成
- 在ShapeNet上生成质量优于现有方法
- 训练推理成本更低，适合高效生成场景
流匹配模型近年来成为连续生成建模的强大框架，适用于3D点云合成。然而，其部署受限于推理时需多步序列采样。均值流（MeanFlow）支持单步生成，显著加速推理，但常难以逼近原始多步流的轨迹，导致样本质量下降。本文提出一种基于最优传输的均值流框架（OT-MF3D），用于高效且准确的3D点云生成与补全。通过引入基于最优传输的采样策略，该方法在保持单步推理的同时，更好地保留了原始多步流的几何与分布结构。在ShapeNet数据集上的实验表明，相比近期基线方法，生成与补全质量均有提升，同时训练与推理成本低于传统扩散和流模型。
原文摘要 · Abstract (English)
Flow-matching models have recently emerged as a powerful framework for continuous generative modeling, including 3D point cloud synthesis. However, their deployment is limited by the need for multiple sequential sampling steps at inference time. MeanFlow enables single-step generation and significantly accelerates inference, but often struggles to approximate the trajectories of the original multi-step flow, leading to degraded sample quality. In this work, we propose an Optimal Transport-enhanced MeanFlow framework (OT-MF3D) for efficient and accurate 3D point cloud generation and completion. By incorporating optimal transport-based sampling, our method better preserves the geometric and distributional structure of the underlying multi-step flow while retaining single-step inference. Experiments on ShapeNet show improved generation and completion quality compared to recent baselines, while reducing training and inference costs relative to conventional diffusion and flow-based models.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。