EVENT DOSSIER
SeamFlow is a new generation generative framework designed to solve the problems of 3D surface cutting and UV unwrapping in computer graphics. This study transforms discrete grid cuts into continuous flow matching tasks in high-dimensional edge probability spaces. Through continuous relaxation learning, the method maps from Gaussian prior distributions to target seam probability distributions, utilizing evolutionary networks to couple local topological markings with global shape priors, and using ordinary differential equations to guide the smoothing of probability flows. Compared with existing autoregressive generative frameworks, SeamFlow uses an edge labeling strategy to improve topological perception while eliminating 3D space projection errors and artificial sequential biases. Experimental results show that this method performs excellently in terms of semantic consistency and parameterization distortion control.
SeamFlow 提出一种新型生成框架，将离散网格切割问题重构为高维边概率空间的连续流匹配，以解决计算机图形学中 3D 表面切割与 UV 展开问题。该框架通过连续松弛学习从高斯先验到目标接缝概率分布的确定性映射，利用演化网络耦合局部拓扑标记与全局形状先验，并通过常微分方程求解引导平滑概率流。相比现有自回归生成框架，SeamFlow 通过边标记化提升拓扑感知能力，同时消除了 3D 空间投影误差和人工顺序偏差。大量实验表明，SeamFlow 实现了卓越的语义一致性和极低的参数化失真。项目页面为 https://meshy-dev.github.io/seamflow。