PaperlayerView exact revision on arXiv

Source-linked AI summary

PartUV: Part-Based UV Unwrapping of 3D Meshes

Zhaoning Wang, Xinyue Wei, Ruoxi Shi, Xiaoshuai Zhang, Hao Su, Minghua Liu

arXiv:2511.16659v2cs.CVcs.CGcs.GR

TL;DR

Existing UV methods can struggle with noisy, poorly conditioned AI-generated meshes, producing fragmented charts and suboptimal boundaries. PartUV combines semantic part decomposition with geometric heuristics in a recursive pipeline. Across diverse datasets, it produces fewer charts and shorter seams while maintaining comparable distortion, with high success on challenging meshes.

  • Problem

    Existing UV unwrapping methods may time out or produce fragmented atlases on bumpy, poorly conditioned AI-generated meshes, complicating texture editing and causing boundary artifacts.

    Source: §1 · p. 2

  • Method

    PartUV combines PartField’s hierarchical semantic decomposition with geometric heuristics in a top-down search that minimizes chart count while keeping distortion below a user-specified threshold.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    Across four datasets, PartUV produces significantly fewer charts and shorter seams while maintaining angular and area distortion comparable to baseline methods, and handles varied meshes with high success.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Semantic part priors align charts with object parts and support organizing related charts across separate atlases.

    Source: §1 · p. 2

    PartUV cannot handle self-intersecting meshes, and meshes with over 1000 components can yield highly fragmented results that may require remeshing.

    Source: §5 · p. 10

Abstract

from arXiv · show

UV unwrapping flattens 3D surfaces to 2D with minimal distortion, often requiring the complex surface to be decomposed into multiple charts. Although extensively studied, existing UV unwrapping methods frequently struggle with AI-generated meshes, which are typically noisy, bumpy, and poorly conditioned. These methods often produce highly fragmented charts and suboptimal boundaries, introducing artifacts and hindering downstream tasks. We introduce PartUV, a part-based UV unwrapping pipeline that generates significantly fewer, part-aligned charts while maintaining low distortion. Built on top of a recent learning-based part decomposition method PartField, PartUV combines high-level semantic part decomposition with novel geometric heuristics in a top-down recursive framework. It ensures each chart's distortion remains below a user-specified threshold while minimizing the total number of charts. The pipeline integrates and extends parameterization and packing algorithms, incorporates dedicated handling of non-manifold and degenerate meshes, and is extensively parallelized for efficiency. Evaluated across four diverse datasets, including man-made, CAD, AI-generated, and Common Shapes, PartUV outperforms existing tools and recent neural methods in chart count and seam length, achieves comparable distortion, exhibits high success rates on challenging meshes, and enables new applications like part-specific multi-tiles packing. Our project page is at https://www.zhaoningwang.com/PartUV.

1 Introduction

Existing UV unwrapping methods can fragment difficult AI-generated meshes and produce boundaries that divide meaningful parts. PartUV combines semantic decomposition with geometric refinement to produce fewer, part-aligned charts with low distortion, and compares favorably with existing tools and neural methods.

1 Introduction

  • AI-generated meshes can cause existing methods to time out or produce extremely fragmented atlases that hinder texture editing and introduce boundary artifacts.Such meshes often have bumpy surfaces, many small triangles, and poor geometric quality, including disconnected components or holes.Source: §1 · p. 2
  • PartUV combines PartField’s hierarchical semantic decomposition with geometric heuristics in a top-down recursive search that minimizes chart count under a user-specified distortion threshold.It uses established parameterization and packing algorithms and includes parallelization and handling for non-manifold and degenerate meshes.Source: §1 · p. 2
  • Semantic priors preserve coherent object parts, support grouping related charts across atlases, and guide seams toward perceptually unobtrusive locations.These properties support editing and rendering tasks and improve the visual quality of textured models.Source: §1 · p. 2
  • Across four datasets, PartUV generates significantly fewer charts and shorter seams while maintaining angular and area distortion comparable to baseline methods.The evaluation includes man-made, AI-generated, CAD, and common 3D shapes.Source: §1 · p. 2
  • PartUV maintains a high success rate on a wide range of meshes, including large, complex, and non-manifold ones, typically processing each mesh in a few to several tens of seconds.The evaluation compares PartUV with xatlas [Young 2019], Blender [Community [n.d.]], Open3D [Zhou et al. 2018a], and recent neural-based methods [Srinivasan et al. 2024].Source: §1 · pp. 2–3

2 Related Work

Prior work addresses UV parameterization, chart segmentation, and learning-based methods through different strategies and trade-offs. PartUV combines semantic part priors with local geometric decomposition in a top-down recursive search.

2.1 Mesh Parameterization

  • Angle-Based Flattening optimizes triangle angles before converting them to UVs and is adopted by PartUV for base unfolding, offering better angle preservation than LSCM at higher computational cost.Parameterization maps a mesh to a 2D domain while minimizing distortion metrics.Source: §2.1 · p. 3

2.2 Mesh Segmentation for UV Unwrapping

  • PartUV combines high-level semantic part priors with local geometric decomposition strategies through a top-down recursive tree search.Existing segmentation approaches include top-down subdivision, bottom-up aggregation, cut optimization, and clustering; common tools primarily use bottom-up aggregation based on local geometry.Source: §2.2 · p. 3

2.3 Learning-Based Parameterization and Segmentation

  • PartUV interleaves semantic segmentation and geometric flattening, addressing the runtime cost of per-shape neural optimization and the limitations of naïvely applying semantic priors to UV decomposition.PartField provides fast, hierarchical 3D part decomposition, while naïve integration can yield suboptimal results.Source: §2.3 · p. 3

3 Method

PartUV combines PartField’s semantic hierarchy with geometric charting and recursive search. It seeks compact chart decompositions under per-chart distortion constraints, with additional processing for efficiency and challenging meshes.

3.1 Overview

  • PartUV combines PartField’s semantic parts with geometric heuristics, recursively seeking few charts that satisfy a user-specified distortion threshold.Charts are flattened with ABF, and the overview also describes checking chart distortion before accepting mappings.Source: §3.1 · pp. 3–4

3.2 Preliminary: PartField

  • PartField predicts a continuous part-feature field; clustering sampled face features builds a hierarchy from individual-face leaves to a whole-mesh root.Similar point features are more likely to belong to the same part.Source: §3.2 · p. 4, §3.3 · p. 4

3.3 Top-Down Recursive Tree Search

  • If primary candidates fail the distortion threshold, the search recurses into PartField children; otherwise, secondary refinement and budgeted subtree search seek a better valid solution.The selected chart set must also satisfy overlap and chart-budget constraints.Source: §3.3 · p. 5

3.4 Geometry-Based Part Decomposition

  • Normal groups connected faces by similar normals; Merge iteratively combines adjacent components when ABF meets distortion and overlap constraints.Normal generates candidate decompositions, while the more expensive Merge heuristic is invoked only when Normal returns an admissible decomposition.Source: §3.4 · pp. 5–6

3.5 Runtime Optimization and Pre- and Post-processing

  • Parallel subtree search and boundary-preserving GPU simplification speed intermediate distortion checks; non-manifold edges are split, and semantic groups support structured multi-atlas packing.Disconnected components trigger recursive exploration, with per-component heuristic decomposition and flattening as a fallback at a predefined depth.Source: §3.5 · p. 6

4 Experiments

PartUV is evaluated across UV-unwrapping baselines, applications, ablations, and worst-case distortion metrics. The results report fewer charts and shorter seams alongside low distortion, and examine how design choices affect performance.

4.1 Implementation Details and Evaluation Setup

  • Experiments use a distortion threshold of 1.25, fixed parameters, and comparisons with Blender, xatlas, Open3D, Nuvo, and OptCuts.PartField is trained on Objaverse; UVPackmaster is used for final group-based packing.Source: §4.1 · p. 6, §4.2 · p. 7

4.2 Comparison with Baselines

  • On Common Shapes, PartUV uses only 1/31 as many charts as Blender, with shorter seams, low angular and area distortion, and runtime comparable to xatlas.The comparisons also report that chart boundaries align more closely with object semantics.Source: §4.2 · p. 8, Table 1 · p. 8
  • Open3D completes only 39.5% of challenging Trellis shapes; on easier completed cases, its chart counts and seam lengths are similar, but its distortion is higher.The reported Open3D averages cover only shapes it successfully processes.Source: §4.2 · p. 8, Table 2 · p. 8
  • Nuvo and OptCuts often take over 30 minutes or hours per mesh; Nuvo has substantially higher distortion, while OptCuts returns outputs for only 9 of 24 Common Shapes.Neither method incorporates semantic parts during optimization.Source: Table 3 · p. 8, §4.2 · p. 9

4.3 Applications and Analysis

  • PartUV’s less-fragmented UV maps support cleaner 2D texture editing and replacement, avoiding artifacts seen with Blender and xatlas.The pipeline also supports packing semantic parts across multiple atlases and lets users adjust the distortion threshold to trade chart count against distortion.Source: §4.3 · p. 9, Figure 7 · p. 9, Figure 8 · p. 9
  • At 128-pixel texture resolution, xatlas and Blender show noticeable color bleeding, while PartUV does not; PartUV also supports semantic multi-atlas packing and threshold adjustment.The 128-pixel setting is described as common in mobile games.Source: §4.3 · p. 9, Figure 10 · p. 10, Figure 9 · p. 10

4.4 Ablation Studies

  • A fixed 20-part PartField decomposition can leave charts too complex for low-distortion flattening; replacing features with face normals doubles runtime and loses semantic alignment.Removing Merge or continued search increases chart count, while removing the distortion surrogate increases runtime.Source: §4.4 · pp. 9–10, Table 4 · p. 10

4.5 95th Percentile Distortion Metrics

  • PartUV’s 95th-percentile area distortion reaches a maximum of 1.442 across four datasets, while baseline values are often higher, including 4.701, 1.885, and 1.728.The evaluation reports both 95th-shape and 95th-chart distortion metrics.Source: §4.5 · p. 10, Table 5 · p. 10

5 Discussion on Failure Cases

PartUV’s performance is constrained by poor input topology, especially self-intersections and meshes with very many components.

5 Discussion on Failure Cases

  • Self-intersecting meshes can trigger deep recursion and fragmented charts; meshes with over 1,000 components can also produce highly fragmented results.The paper notes remeshing may be required for extremely fragmented inputs.Source: §5 · p. 10
  • PartUV performs well on general meshes with few components and does not require inputs to be watertight or manifold.Source: §5 · p. 10
  • Poor input mesh topology is the stated boundary on robustness; self-intersections are one specific challenging case.Source: §5 · p. 10

6 Conclusion

PartUV combines learning-based semantic part priors with geometric heuristics for UV unwrapping. The paper concludes that it produces fewer charts, low distortion, and boundaries aligned with semantic parts, and demonstrates applications of the pipeline.

6 Conclusion

  • PartUV integrates semantic part priors from learning-based methods with two geometric heuristics.Source: §6 · p. 10
  • The framework generates significantly fewer charts, maintains low distortion, and aligns chart boundaries with semantic parts.Source: §6 · p. 10
  • The paper demonstrates the pipeline’s advantages through several applications.Source: §6 · p. 10

Appendix

The appendix materials include a visualization of distortion effects and an efficiency comparison across methods.

Appendix

  • Figure 12 illustrates curved stripes from angular distortion and uneven stripes from area distortion.Source: Figure 12 · p. 11
  • Table 6 presents an efficiency comparison across methods.Source: Table 6 · p. 11

A1.1 More Qualitative Results

The supplementary material provides additional qualitative and quantitative comparisons across the four datasets, with metrics reported for individual shapes.

A1.2 Efficiency

PartUV remains competitive with baseline methods in UV-map efficiency, measured by the total area of valid 2D faces in normalized UV space.

  • PartUV does not hurt overall UV-map efficiency and remains competitive with baseline methods.Efficiency is the total area of valid 2D faces within the normalized 0–1 UV space.Source: A1.2 Efficiency · p. 11
  • Charts from the same parts are always grouped during packing, using UVPackMaster to create the final packed UV map.Source: A1.2 Efficiency · p. 11
  • UVPackMaster’s heuristic search was set to 3 seconds, although most meshes finished within microseconds.Source: A1.2 Efficiency · p. 11

A1.3 Analysis and Visualization of Open3D Results

Open3D frequently crashed or exceeded its time budget on challenging meshes, and successful outputs could still show conspicuous distortion and geometry-disregarding seams.

A2 Additional Ablation Results

The ablations show that PartField features organize charts better than face normals, while the Normal and Merge heuristics complement one another and ABF++ improves Trellis results over LSCM.

A2.1 Replacing PartField feature with Face Normals

A2.2 Without Merge Heuristic

A2.3 Without Normal Heuristic

  • Using Merge alone raises runtime on bumpy meshes; on bimba, runtime rises from 147 seconds to over 892 seconds.Bumpy geometry can cause overlapping during merging and increase chart count, so Merge is most effective as a complement to Normal on sufficiently good results.Source: A2.3 Without Normal Heuristic · p. 12, Figure 15 · p. 12

A2.4 Using LSCM instead of ABF++

Related papers

Loading 2511.16659v2…