SIGGRAPH Asia 2026 · Conference Papers

SeamFlow Structure-Aware Flow Matching on Edge Probabilities for Artist-Like UV Unwrapping

Yuming Zhao*1 Zangyueyang Xian*2 Qijian Zhang3 Rendong Liang2 Qin Jia2 Ying He4 Junhui Hou†1
1 City University of Hong Kong 2 Meshy AI 3 Bambu Lab 4 Nanyang Technological University * Equal contribution · † Corresponding author
Paper DOI Code soon Models soon
SeamFlow teaser showing diverse meshes and a mesh seam layout evolving from Gaussian noise into a coherent cut graph.

From Gaussian edge noise to coherent seam graphs: SeamFlow generates artist-like cut layouts across diverse objects while staying precisely aligned with mesh topology.

Overview

Abstract

UV cutting and unwrapping are fundamental to computer graphics, but optimizing only geometric distortion can ignore the semantic coherence expected from artist-authored layouts. Existing autoregressive generators improve semantics, yet can introduce projection errors, artificial sequence-order bias, and locally inaccurate cuts.

SeamFlow instead relaxes discrete mesh cutting into continuous flow matching over a high-dimensional edge-probability space. An edge-token evolution network combines local mesh topology with global shape context and integrates a probability-flow ODE from Gaussian noise to a seam distribution. The result is a topology-aware, order-free generator that produces semantically coherent seams with low distortion.

Core contributions

Structure-aware generation on mesh edges

01

Topology-aware

Edge tokens explicitly encode mesh connectivity, keeping predicted cuts aligned with the underlying surface without 3D projection.

02

Flow-matched

All edge probabilities evolve in parallel through a continuous flow, removing artificial autoregressive order and sequential decoding.

03

Globally conditioned

Local topological tokens are fused with a global shape prior to preserve long-range symmetry, components, and semantic structure.

Method

From edge noise to UV charts

SeamFlow predicts the probability of cutting every mesh edge in one continuous process, then converts the resulting seam graph into a UV map with standard parameterization.

SeamFlow pipeline with a global encoder, structure-aware flow matching over edge tokens, predicted edge probabilities, and a UV map.
Pipeline overview. Global geometric features condition a transformer that denoises edge tokens into precise seam probabilities.
1

Encode the shape

A global encoder extracts semantic context from coordinates and surface normals.

2

Evolve edge tokens

Self- and cross-attention transport noisy edge probabilities along the learned flow.

3

Unwrap the mesh

The final seam probabilities define a cut graph that is parameterized into UV islands.

Results

Coherent, diverse, and scalable

Evaluation on 2,000 unseen test meshes covers geometric distortion, seam rationality, runtime, user preference, texture generation, editability, and cross-resolution generalization.

0.12Angle distortion ↓
0.14Area distortion ↓
5.63 sRuntime ↓
61.0%User study Top-1
81.8%User study Top-2

Selected values from the paper's evaluation setup; distortion and runtime are lower-is-better.

Surface cutting · Fig. 8
Qualitative comparison of SeamFlow and baseline UV unwrapping methods on a complex fire extinguisher model.
Complex-model comparison. SeamFlow places fine-grained cuts around dense geometric details while keeping long, compact seams in simpler regions. Baselines tend to over-segment the surface or miss important local structure.
Downstream texture · Fig. 9
Texture generation comparison on a fire extinguisher, showing that SeamFlow preserves sharp text and logo continuity across seams.
3D texture generation. Under identical text prompts, SeamFlow's coherent UV islands preserve sharp continuity in the “FIRE” lettering and safety icons. Competing layouts introduce visible blur, duplication, and discontinuities.

Complex geometry

Fine cuts adapt to handles, joints, and elongated branches, while broad smooth regions retain compact long-path seams.

Texture continuity

Semantically aligned islands reduce discontinuities across object boundaries and preserve high-frequency text and logo details.

User preference

In the paper's user study, SeamFlow results were selected in the Top-1 by 61.0% and in the Top-2 by 81.8% of participants.

Reference

BibTeX

@inproceedings{zhao2026seamflow,
  author    = {Zhao, Yuming and Xian, Zangyueyang and Zhang, Qijian and
               Liang, Rendong and Jia, Qin and He, Ying and Hou, Junhui},
  title     = {SeamFlow: Structure-Aware Flow Matching on Edge Probabilities
               for Artist-Like UV Unwrapping},
  booktitle = {SIGGRAPH Asia 2026 Conference Papers},
  year      = {2026},
  articleno = {111},
  numpages  = {12},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  doi       = {10.1145/3829340.3842329},
  url       = {https://doi.org/10.1145/3829340.3842329}
}