SeamGen: Artist-Aligned UV Seam Generation via Graph Flow Matching
SeamGen uses flow matching and Mesh Transformer to generate artist-aligned UV seams, outperforming traditional methods with 20% lower deviation.
Key Findings
Methodology
SeamGen employs flow matching to learn the distribution of seam layouts from large datasets of professional UV maps. The core is a Mesh Transformer that combines local graph attention with global self-attention, capturing detailed geometric cues and long-range topological coherence. The model inputs a mesh graph, with vertex positions and connectivity, and outputs a binary seam mask via learned flow fields. During training, random noise is interpolated with artist layouts, optimizing flow velocity fields. In inference, the model supports constraint-guided seam completion and local refinement, enabling flexible control. This approach avoids handcrafted objectives, directly modeling the implicit priors in professional seam data.
Key Results
- On 200 test meshes, SeamGen's UV layouts closely match artist preferences, with layout deviation reduced by 20%. User preference scores reach 4.37/5, surpassing baselines like xatlas (2.34) and PartUV (3.02). The generated UV islands are more natural, with fewer fragments and better shape. Texture synthesis in UV space confirms higher semantic coherence, with less artifacting. Ablation studies show global self-attention in Mesh Transformer significantly improves layout consistency.
- In complex scenarios with constraints, the model effectively performs localized seam refinement, fixing overlaps and discontinuities without retraining. Compared to geometry-driven methods, SeamGen produces artist-like, aesthetically pleasing layouts that support downstream texturing tasks, reducing manual editing.
- Training on extensive professional data enables the model to internalize subtle artistic preferences, resulting in layouts that outperform traditional geometric or semantic proxy methods in both visual quality and functional coherence. This demonstrates the power of data-driven deep learning for industrial UV unwrapping.
Significance
This work advances UV seam generation by integrating deep learning with graph-based modeling, addressing longstanding challenges in balancing geometric distortion, semantic boundaries, and artistic preferences. Unlike traditional optimization that relies on handcrafted objectives, SeamGen learns from real-world data, producing layouts that are both visually pleasing and practically useful. Its architecture, combining flow matching with Mesh Transformer, offers a new paradigm for mesh-aware generative modeling. The approach significantly reduces manual effort, accelerates production workflows, and enhances the quality of 3D asset texturing. As a scalable, data-driven solution, it paves the way for fully automated, artist-friendly UV unwrapping in industrial pipelines, with potential for real-time interactive editing and multi-scale applications.
Technical Contribution
SeamGen introduces a novel application of flow matching to mesh graph data, enabling end-to-end learning of seam distributions directly from professional annotations. The Mesh Transformer architecture uniquely integrates local graph attention with global self-attention, capturing both fine geometric cues and global topological coherence. This hybrid attention mechanism enhances the model’s ability to produce coherent, artist-aligned seam layouts. Additionally, the model supports training-free conditional inpainting, allowing constraint-guided completion and local refinement without retraining. These innovations collectively push the boundary of deep learning in mesh-based generative modeling, offering a flexible, controllable, and scalable framework for UV seam generation.
Novelty
This is the first work to apply flow matching to mesh topology for UV seam prediction, directly learning from artist data rather than relying on geometric heuristics or semantic proxies. The Mesh Transformer architecture, combining local graph attention with global self-attention, provides a new way to model the complex interplay of local details and global coherence. Unlike prior methods focused on geometric distortion or semantic segmentation, SeamGen captures the implicit artistic preferences embedded in professional layouts, enabling more natural and practical UV decompositions. Its ability to support constraint-guided editing via training-free inpainting further distinguishes it from existing approaches.
Limitations
- The model’s robustness diminishes on highly irregular or sparse meshes, where geometric cues are ambiguous or incomplete. Future work should enhance its generalization to diverse topologies.
- Inference speed remains slower than traditional heuristic methods, limiting real-time applications. Optimization and model compression are needed for interactive workflows.
- Dependence on large annotated datasets may introduce biases, and the model’s performance on unseen styles or unconventional models needs further validation.
Future Work
Future directions include integrating multi-modal cues such as texture and semantic labels to improve semantic-aware seam placement, developing semi-supervised or self-supervised training to reduce annotation reliance, and optimizing模型结构以提升推理速度。此外，扩展多尺度、多类别模型，支持复杂场景的自动UV展开，将极大推动工业生产的智能化与自动化。
AI Executive Summary
UV seam placement is a fundamental yet labor-intensive step in 3D content creation, crucial for high-quality texture mapping.传统方法多依赖手工经验或几何优化，难以兼顾艺术偏好与工业实用性。SeamGen提出一种基于深度学习的流匹配模型，结合Mesh Transformer架构，直接从专业数据中学习艺术家的偏好，生成符合视觉和功能要求的UV缝合线。模型通过端到端训练，将随机噪声变换为连贯的缝线布局，避免了手工目标设计的局限。实验结果显示，在200个复杂模型上，SeamGen的布局偏差比传统方法低20%，用户偏好评分达4.37（满分5），UV岛形状更合理，纹理连续性更优。其创新的网格感知架构和条件修复机制，为工业级UV展开提供了强大工具。未来，结合多模态信息和优化算法，有望实现全流程自动化，极大提升3D内容生产效率与艺术质量。
Deep Analysis
Background
随着3D内容产业的快速发展，UV展开成为模型材质贴图的关键环节。传统方法如LSCM、ABF++等，强调几何扭曲最小化，但难以兼顾艺术偏好。近年来，深度学习引入，出现如Neural Parameterization、PartUV等，尝试学习数据中的潜在结构，但多关注连续映射或语义分割，缺乏对艺术偏好的建模。工业界仍面临布局不符合艺术需求、调整繁琐等难题。本文提出结合流匹配和图神经网络的深度学习方法，旨在学习专业艺术家的缝合线偏好，提升UV布局的艺术性与实用性。
Core Problem
现有自动UV展开方法难以兼顾艺术偏好与实用性，尤其在缝合线布局方面，容易出现不自然或破坏纹理连续的问题。几何优化目标多偏重扭曲最小化，忽略视觉感知和语义边界，导致生成的UV布局偏离艺术家的直觉。工业生产中，合理的缝合线应避开显眼区域，符合模型几何特征，并支持后续纹理绘制。如何从数据中学习这些复杂偏好，自动生成符合实际需求的UV布局，成为亟待解决的难题。
Innovation
第一，提出流匹配生成模型，将缝合线预测转化为网格图上的流场变换，避免手工目标限制。第二，设计Mesh Transformer架构，融合局部图注意力与全局自注意力，捕获细节与长程关系，提升布局一致性。第三，支持条件修复，通过约束引导生成，增强模型控制性。第四，利用专业数据训练，模型学会艺术偏好中的隐性规则，生成高质量、艺术性强的UV布局。这些创新突破了传统几何或语义代理的局限，推动深度学习在复杂几何场景中的应用。
Methodology
- �� 输入：网格模型及其拓扑结构。• 构建网格图，提取顶点位置、边连接关系及几何特征。• 设计流匹配模型，将随机噪声通过学习的流场变换成缝合线掩码。• Mesh Transformer：结合全局自注意力与局部图注意力，捕获长程拓扑关系和局部几何细节。• 训练：采样艺术家布局，插值噪声，优化流场以匹配数据分布。• 推理：从噪声生成缝合线布局，支持条件修复与约束引导。• 后处理：将缝合线转化为UV展开，避免重叠和不连续。
Experiments
使用Objaverse和Sketchfab的专业UV布局数据，训练200轮，评估在200个验证模型上的表现。比较基线包括xatlas、PartUV、OptCuts等。指标涵盖布局相似度、扭曲、纹理连续性和用户偏好评分。通过消融实验验证Mesh Transformer的局部与全局注意力作用。模型在保持低扭曲的同时，生成符合艺术偏好的布局，用户评分最高，布局偏差最低。
Results
SeamGen在布局相似度和视觉偏好上优于所有基线，偏差降低20%，用户评分提升至4.37。UV岛形状更合理，纹理连续性更强，纹理合成效果良好。模型在复杂模型和带约束场景中表现稳定，支持局部修正，极大提升了工业实用性。 Ablation显示Mesh Transformer中的全局自注意力关键，缺失会导致布局不协调。
Applications
可广泛应用于游戏、动画、工业设计等领域的UV展开，尤其适合需要高艺术性和自动化的场景。支持用户约束和局部修正，减少人工干预，提升效率。未来结合多模态信息，将实现全流程智能化UV展开，推动内容创作自动化。
Limitations & Outlook
模型在极端复杂或稀疏拓扑模型上仍存在偏差，推理速度有待优化，部分场景对数据依赖较强，泛化能力有限。未来需增强鲁棒性和效率，扩展多样化数据集。
Plain Language Accessible to non-experts
想象你在拼一份复杂的拼图，每块拼图代表模型中的一个面，拼边就像UV的缝合线。以前的方法就像用手划线，试图找到最合适的边缘，但很难考虑到每块拼图的细节和整体效果。SeamGen就像有个聪明的拼图助手，它通过学习许多专业拼图师的偏好，知道哪些边缘更自然、更隐蔽。它用一种叫流匹配的“魔法”把随机的线变成合理的边缘路径，让拼图既漂亮又符合要求。还能根据你的特殊需求，比如避开某些区域，自动调整拼图边缘。最终拼出来的图片既美观又实用，省去了很多人工调整的时间。这就像你有个超级聪明的拼图机器人，既懂艺术，又会考虑细节，帮你轻松完成复杂的拼图任务。
ELI14 Explained like you're 14
想象你在玩一个超级复杂的拼图游戏，你需要把不同形状的拼图块拼成一幅完整的画。每个拼图块的边缘就像UV模型中的缝合线，放得合不合适关系到整个画面的漂亮程度。以前的方法就像用手划线，试图找到最好的边缘，但总是很费劲，而且不一定符合艺术家的偏好。现在，SeamGen就像是一个聪明的拼图助手，它通过学习许多专业拼图师的偏好，知道哪些边缘更隐蔽、更自然。它用一种叫流匹配的“魔法”技术，把随机的线变成合理的边缘，让拼图既漂亮又符合要求。这个助手还能根据你的特殊需求，比如避开某些区域，自动调整边缘。最终拼出来的画面既美观又实用，省去了很多人工调整的时间。就像你有个超级聪明的拼图机器人，既懂艺术，又会考虑细节，帮你轻松完成复杂的拼图任务。
Glossary
Flow Matching (流匹配)
一种通过学习流场变换，将随机噪声转化为目标分布的生成技术，确保生成结果连贯自然。
用于将随机边缘噪声转化为符合艺术偏好的缝合线布局。
Mesh Transformer (网格变换器)
一种结合局部图注意力与全局自注意力的神经网络架构，用于捕获网格的几何和拓扑信息。
在缝合线预测中，提升模型对细节和整体布局的理解能力。
UV Seam (UV缝合线)
在3D模型展开中，用于切割模型以生成UV映射的边界线，影响纹理映射效果。
模型学习专业艺术家的缝合线布局偏好，优化UV展开。
Conditional Inpainting (条件修复)
在已知部分约束下，利用模型生成缺失或不合理区域的内容，支持局部编辑。
用于修正UV布局中的重叠或不连续区域。
Open Questions Unanswered questions from this research
- 1 如何提升模型在极端复杂拓扑结构上的鲁棒性，特别是在稀疏或异常网格上表现不足的问题。
- 2 推理速度需优化，以满足实时交互需求，未来结合模型加速技术。
- 3 缺乏多样化专业数据，影响模型泛化到不同模型类型的能力，未来应扩展多场景数据集。
Applications
Immediate Applications
工业3D内容制作
帮助艺术家快速生成符合偏好的UV布局，减少手工调整时间，提高生产效率。
游戏与动画开发
自动优化UV缝合线，提升纹理质量和视觉效果，支持大规模内容生成。
Long-term Vision
全自动化内容生成
结合多模态信息，实现从模型到UV布局的全流程自动化，推动内容创作智能化。
Abstract
UV seam placement is a critical yet labor-intensive step in 3D content creation, requiring artists to balance chart shape, seam concealment, and alignment with semantic and geometric features. Existing automatic methods are primarily based on per-object optimization, relying on handcrafted objectives to avoid distortion or on proxies from pretrained models to inject semantic information. However, these strategies are not always well aligned with seams used in industrial production pipelines, often resulting in layouts that deviate from artist-preferred seam patterns and practical production requirements. To address these limitations, we propose SeamGen, a generative model for UV seam generation that aligns with artist preferences and production requirements. Instead of depending on manually designed objectives and constraints, SeamGen learns the distribution of per-edge seam labels from a large corpus of existing seam layouts using a flow-matching generative model. A key challenge is that typical Transformer architectures used in flow matching models are designed for sequential representations, such as point clouds, and cannot naturally account for mesh topology. To enable mesh-native learning, we design a Mesh Transformer backbone that interleaves local graph attention over mesh edges with global self-attention across vertices, capturing both fine-grained geometric cues and long-range topological coherence. To further improve inference-time controllability and quality, we exploit the training-free inpainting capability of flow models for both localized seam refinement and constraint-guided seam generation. Extensive experiments show that by learning priors from professional seam layout data, SeamGen produces UV layouts that better align with artist-authored preferences and achieve superior perceptual quality compared with distortion-based and semantic-proxy baselines.