PaperlayerView exact revision on arXiv

Source-linked AI summary

MeshQuery: Agentic Seam Planning for UV Parametrization

Marco Schouten, Arthur Roullier, Elie Michel, Ruben Wiersma, Axel Paris, Tamy Boubekeur

arXiv:2610.02507v1cs.CV

TL;DR

Automated UV seam planning must reconcile artistic intent with geometric constraints, while existing approaches have difficulty covering cases or scaling. MeshQuery uses a VLM with queryable mesh data, a seam-selection DSL, and iterative UV feedback; it produces fewer charts and shorter seams than PARTUV, with professional artists preferring its results in most comparisons.

  • Problem

    Automatic seam placement must balance semantic and downstream-use considerations with geometry, but existing methods do not handle all cases well and data-driven approaches struggle to scale and require difficult-to-acquire high-quality seam data.

    Source: §1 · p. 2

  • Method

    MeshQuery has a VLM query mesh attributes on demand, express seam plans as DSL programs of edge-selection operators, and iteratively refine plans using UV-quality feedback.

    Source: §3 · p. 3, §3.1 · p. 4

  • Results

    Against PARTUV, MeshQuery yields 2.90×/4.29× fewer UV charts and 1.63×/1.69× shorter UV seams on SUBSTANCE-50/TOYS4K-50; professional artists prefer it in 80.9% of comparisons.

    Source: §4.2 · p. 6

  • Takeaways & Limitations

    The compact mesh and DSL representation supports VLM reasoning on large meshes, with MeshQuery scaling to 300k polygons, an order of magnitude beyond autoregressive methods.

    Source: §4.2 · p. 7, §5 · p. 9

    The method focuses on quad meshes, and its runtime has bottlenecks in PartField processing for larger meshes and the iterative agentic loop.

    Source: §5 · p. 9

Abstract

from arXiv · show

We present MeshQuery, a training-free agentic approach to automatic UV unwrapping of production-grade quad meshes. A Vision-Language Model (VLM) plans artist-aligned seams using a set of edge-selection tools, conditioned on domain-specific UV-unwrapping knowledge expressed in natural language and refined with a feedback loop. We design a queryable mesh representation together with a domain-specific language (DSL) that enables the agent to retrieve mesh information on demand, express a seam plan as a compact program of edge-selection operators over topological, geometric, and semantic mesh attributes, and iteratively refine it from UV quality feedback. On Adobe Substance 3D and Toys4K meshes, MeshQuery produces 2.9x/4.29x fewer charts and 1.63x/1.7x shorter seams than the strongest baseline, and professional artists prefer its results in 80.9% of comparisons. Ultimately, decoupling high-level intent planning from low-level edge selection and compact mesh representation lets MeshQuery run on different backend VLMs and scale to meshes an order of magnitude larger than autoregressive seam prediction

1 INTRODUCTION

UV seam planning must balance high-level considerations such as semantics and downstream use with low-level geometry, yet existing automated methods do not handle all cases well. MeshQuery uses a VLM, a compact queryable mesh representation, and a seam-selection DSL to plan and refine seams on quad meshes.

1 INTRODUCTION

  • Existing automatic seam methods handle cases unevenly, while recent data-driven approaches struggle to scale and require difficult-to-acquire high-quality seam training data.As a result, UV unwrapping remains manual and time-consuming.Source: §1 · p. 2
  • MeshQuery combines VLM planning with a mesh-query interface and seam-selection DSL, grounding high-level reasoning in precise edge selection.Its DSL combines geometric and connectivity attributes, while active perception requests mesh information selectively.Source: §1 · p. 2
  • MeshQuery targets quad meshes and reports results on organic and hard-surface assets, scaling to meshes with more than 300k faces.Its contributions include an active-perception mesh representation and compact DSL, alongside prompt refinement and quantitative and user-study evaluation.Source: §1 · p. 2

2 RELATED WORK

Prior seam-planning approaches optimize geometric objectives, learn from seam examples, or adapt general agent interfaces to 3D. MeshQuery instead uses a queryable mesh interface and an agentic planning loop to address high-level seam intent with tool-based selection.

2 RELATED WORK

  • Optimization-driven methods target seam length and distortion, but their objectives typically omit artistic considerations such as semantics.Approaches include heuristic cutting, differentiable objectives, neural representations, and specialized optimizers.Source: §2 · p. 2
  • Data-driven seam methods rely on high-quality seam examples, and graph-network and autoregressive approaches struggle with large meshes.MeshQuery instead uses VLM knowledge and low-level tools in an agentic feedback loop, scaling to meshes with more than 300K vertices.Source: §2 · p. 3
  • Mesh-token serialization is costly for high-resolution meshes, while latent projections tie representations to particular VLMs; MeshQuery retrieves task-relevant information on demand.The framework also supports any backend VLM.Source: §2 · p. 3
  • General agent interfaces abstract low-level actions through tools and programs, while MeshQuery uses relational database queries to avoid complex 3D camera placement for point-and-click seam selection.The cited interface work concerns how observation and action representations affect agent performance.Source: §2 · p. 3

3 METHOD

MeshQuery separates seam planning into a VLM-facing mesh query representation, a DSL for edge selection, and a UV-quality feedback loop. A seam-preference skill guides higher-level planning while the agent iteratively refines its plan.

3 METHOD

  • The agent queries mesh attributes, composes a seam-selection program, and evaluates its instantiated plan with UV metrics returned as feedback.It repeats the process until the stopping criterion is met; the seam-preferences skill supplies higher-level direction.Source: §3 · pp. 3–4

3.1 QUERYABLE MESH REPRESENTATION

  • The preprocessor stores mesh attributes in a relational database so SQL queries can retrieve compact information instead of serializing the whole mesh into the VLM context.Attributes include connectivity, geometric measurements, and semantic labels; ambient occlusion serves as a proxy for edge visibility.Source: §3.1 · p. 4

3.2 SEAM-SELECTION DSL

  • The DSL represents a seam plan as a program of edge-selection statements, with dependency ordering and type checks rejecting invalid programs.Operators return aggregate edge sets; preview and search variants help the agent test non-empty selections and find ranked candidates when exact arguments are difficult to specify.Source: §3.2 · p. 4

3.3 UV QUALITY FEEDBACK LOOP

  • UV feedback combines global atlas measures with local chart measures, including distortion, chart convexity, seam connectivity, and flags for high-distortion charts.The agent can request distortion heatmaps, and the loop stops when distortion bounds are met or the time budget is exceeded.Source: §3.3 · p. 5
  • A lexicographic ranking prioritizes feasibility, then fewer sliver charts and lower distortion, followed by convexity and UV-space utilization.The preferences skill can steer this design toward other specific goals.Source: §3.3 · p. 5

3.4 SEAM PREFERENCES SKILL

  • The seam-preferences skill combines tool and DSL guidance with manually distilled seam-placement recommendations for common shapes and situations such as head seams.The recommendations are derived from publicly available handbooks.Source: §3 · p. 4, §3.4 · p. 5

4 EXPERIMENTS

Experiments compare MeshQuery with UV-unwrapping baselines on two 50-asset quad-mesh datasets, then evaluate artist preferences, scaling, and ablations. Results include fewer charts and shorter seams than PARTUV, alongside improved distortion acceptance with feedback.

4.1 EXPERIMENTAL SETUP

  • MeshQuery is evaluated on 50-asset SUBSTANCE-50 and TOYS4K-50 quad-mesh datasets, using chart, seam, visibility, shape, and distortion metrics.The datasets span 1,344–53,634 and 61–330,640 polygons, respectively; the VLM is Claude Opus 5 with a 1M-token context.Source: §4.1 · pp. 5–6

4.2 RESULTS

  • 80.9% of professional artists’ decisive ratings preferred MeshQuery over PARTUV; across all annotators, preference was 79.7%.Ours-preferred seams were marked usable as a production hot-start in 53.0% of cases.Source: §4.2 · p. 6
  • Qualitative comparisons show minimal seam sets, generally placed in occluded regions or along high-dihedral-angle edges, and following natural object structure when attributes are at parity.Figure 4 presents checkerboards and seams above the resulting UV atlases for TOYS4K-50 assets.Source: §4.2 · p. 6, Figure 4 · p. 7
  • Tool-call counts correlate negligibly with polygon count (Spearman ρ = +0.11), while runtime scales super-linearly and MeshQuery reaches 300k polygons.For similar semantics, geometry, and topology, tool calls interpolate toward a constant; runtime is bounded by feedback parameterizations and sparse solves.Source: §4.2 · pp. 6–7, Figure 5 · p. 8
  • Mesh preprocessing takes a median 2.6 minutes, with VLM labeling accounting for 77.9% of that cost.The relational representation takes a median 2.8 seconds; Figure 9 is referenced for preprocessing details.Source: §4.2 · p. 7
  • 2.90×/4.29× fewer UV charts and 1.63×/1.69× shorter UV seams than PARTUV on SUBSTANCE-50/TOYS4K-50, with minimal distortion cost.Table 1 reports the comprehensive benchmark; methods with failures or excessively high mean distortion are excluded from ranking.Source: §4.2 · p. 6, Table 1 · p. 8

4.3 ABLATION STUDIES

  • UV feedback raises the share of meshes within acceptable distortion bounds from 66% on the first attempt to 97% after revisions.Across 100 meshes, the agent applied 342 plan versions; most gains occurred within three iterations, reaching 91 accepted plans.Source: §4.3 · p. 8, Table 2 · p. 9, Figure 6 · p. 9
  • Claude Opus 5 substantially improves benchmark distortion metrics over Qwen3.8-27B, though the smaller open-source model completes the task at higher distortion cost.The comparison highlights portability to smaller models despite the stronger results from frontier models with greater reasoning and token capacity.Source: §4.3 · p. 8, Table 2 · p. 9

5 CONCLUSION

MESHQUERY uses compact mesh queries, a seam-selection DSL, and parameterization feedback to plan UV seams. The conclusion reports strong artist preference while identifying mesh scope and runtime costs as challenges.

5 CONCLUSION

  • Professional artists prefer MESHQUERY results over SOTA baselines in 80.9% of comparisons, and rate 53% of meshes usable as production hot-starts.The evaluation covers small to large meshes and hard and organic surfaces.Source: §5 · p. 9
  • MESHQUERY combines compact mesh queries, a seam-selection DSL, and parameterization feedback to refine seam plans.Aggregate operators let the VLM plan at a higher level than individual edge selection.Source: §5 · p. 9
  • The framework focuses on quad meshes, although the authors say its DSL, mesh queries, and skill files theoretically generalize to triangular meshes.Early explorations with SeamCut are described as encouraging.Source: §5 · p. 9
  • PartField preprocessing and the iterative agentic loop are runtime bottlenecks; future work also targets joint seam placement and UV placement.The authors expect more efficient agent frameworks and more informative evaluation signals to reduce loop cycles.Source: §5 · p. 9

A.1 DATASETS

The appendix describes two evaluation datasets and the tools exposed through the agent interface. SUBSTANCE-50 contains production-grade artist-composed quad meshes, while TOYS4K-50 covers single-component 2-manifold meshes.

A.1 DATASETS

  • SUBSTANCE-50 spans 18 distinct Adobe Asset Library categories and includes selected production-grade, manually artist-composed quad meshes; TOYS4K-50 covers 44 categories and contains single-component, 2-manifold meshes.Source: §A.1 · p. 14
  • SUBSTANCE-50 covers 18 Adobe Asset Library categories and includes multi-component meshes, whereas TOYS4K-50 covers 44 categories of single-component, 2-manifold meshes.The datasets also differ in their counts of bilaterally symmetric assets and assets with estimated genus of at least two.Source: §A.1 · p. 14

A.2 DSL OPERATORS

  • The agent interface provides one tool for each DSL operator, plus tools for mesh observation and seam-plan management.The appendix reports each tool’s signature, description, and time complexity.Source: §A.2 · p. 14

A.3 MESH REPRESENTATION

The mesh representation stores topology, geometric signals, semantic labels, and precomputed edge candidates in a relational database. Preprocessing also creates connected semantic clusters and normalized features for querying.

A.3 MESH REPRESENTATION

  • The database organizes assets, scenes, meshes, components, faces, edges, adjacency, and precomputed feature candidates, including component topology and edge incidence.Coordinates use a Blender Z-up frame with isotropic normalization; pct denotes a within-mesh rank in [0, 1], while raw retains world units or radians.Source: §A.3 · p. 14
  • PartField segmentation is cut into eight clusters, face labels are transferred to polygons, and Claude Sonnet 5 names the object and labels up to eight clusters.Remaining clusters are unlabeled, and edges between differently labeled faces are flagged as semantic boundaries.Source: §A.3 · p. 14
  • Geometric features include edge-level ambient occlusion, thickness, signed curvature, dihedral angle, normals, radial normal, and normalized edge length.Ambient occlusion uses eight hemisphere rays per edge midpoint; thickness uses four inward rays and is NULL for open sheets.Source: §A.3 · p. 14
  • Precomputed candidates include connected transition-edge rings, curvature ridges, genus necks, and symmetry planes.Ring and ridge candidates have minimum six-edge runs; symmetry tests the three axis-aligned planes through the bounding-box center.Source: §A.3 · pp. 14–15
  • Figure 7 compares dataset mesh-size distributions by vertex count and reports that Substance-50 and Toys4k-50 are comparable to other common 3D mesh datasets.Source: Figure 7 · p. 15

A.4 BASELINES

The appendix specifies baseline settings and a mesh-dependent rule for choosing the parameterization algorithm. Baselines share a 40-minute budget, with timed-out samples counted as failures.

A.4 BASELINES

  • Every baseline receives a 40-minute budget, and samples that time out are recorded as failed.Source: §A.4 · p. 15
  • Baseline configurations vary: Blender Smart UV uses a 66° angle limit and 0.03 island margin, while XATLAS uses default chart options.The appendix also specifies settings and seeds for OptCuts, PARTUV, and FlattenAnything.Source: §A.4 · p. 15

A.5 PARAMETERIZATION ALGORITHM

  • The framework selects LSCM when the share of dihedral-angle edges exceeds a threshold; otherwise it uses Min. Stretch, with a possible fallback to ABF++ if distortion thresholds are unmet.The choice is based on the distribution of dihedral-angle edges.Source: §A.5 · p. 15

A.6 METRICS

The appendix defines metrics for UV distortion, seam placement, and chart quality. It also reports that distortion performance does not degrade as meshes grow larger.

A.6 METRICS

  • Chart count is the number of connected UV components, while non-disk fraction measures the share of charts whose Euler characteristic differs from one.Compactness and convexity assess chart area relative to perimeter and convex-hull area; jaggedness averages second differences along resampled boundary loops.Source: §A.6 · pp. 15, 17, 18
  • 3D seam length is cut-edge length divided by total surface area, while seam visibility compares seam occlusion against occlusion across the mesh.Visibility uses 1,000 evenly spaced seam samples and compares every seam value with every reference value, counting ties as half.Source: §A.6 · pp. 15, 18
  • The metric suite also includes UV seam length, defined as the total UV length of chart-boundary edges.Each physical cut edge is counted once when computing 3D seam length.Source: §A.6 · p. 15
  • Area distortion averages capped, atlas-normalized face-area ratios over kept faces; worst-chart distortion takes the largest chart-specific mean.Figure 10 reports that distortion performance does not degrade over larger meshes.Source: §A.6 · p. 18, Figure 10 · p. 19
  • Angular distortion measures face-axis non-orthogonality, while conformal distortion averages the singular-value ratio over eligible faces.Angular distortion is one for perpendicular UV axes; conformal distortion omits flipped and collapsed triangles, with one ideal and lower better.Source: §A.6 · p. 18
  • Texel spread is the 3D-area-weighted standard deviation of per-face log texel density, measured in octaves.It is zero when all faces receive the same texture resolution and is independent of global UV scale.Source: §A.6 · p. 18

A.7 MEASURING TOKEN COUNT

Tool-call observations are substantially smaller than serialized mesh input at the median, and their size changes much less with mesh resolution.

A.7 MEASURING TOKEN COUNT

  • 36k observation tokens per run was the median across 100 meshes, versus 230k tokens for serialized OBJ meshes.Observation tokens are the text returned by tool calls; OBJ serialization is the mesh-file byte size, converted at four characters per token.Source: §A.7 · p. 19
  • Observation-token counts were nearly flat with polygon count (Spearman ρ = +0.15; slope 0.05 per decade), while serialized size grew linearly (ρ = +0.99; slope 1.06 per decade).Both scaling estimates use log–log fits.Source: §A.7 · p. 19
  • The largest per-turn context, including cached prompt and prior tool results, had a median of 157k tokens and a maximum of 503k.Across 100 meshes, observation-token counts ranged from 10k to 175k.Source: §A.7 · p. 19

B EXTENDED RESULTS

Extended results report distortion scaling over larger meshes and summarize agent tool use across the evaluation sets. The appendix also extends qualitative comparisons and shows the evaluation assets.

B EXTENDED RESULTS

  • Distortion performance was uncorrelated with mesh polygon size, and the method scaled well to larger meshes.Figure 10 extends the main paper's pointwise distortion plots.Source: B EXTENDED RESULTS · p. 19
  • Across 3,049 tool calls, 30.9% were queries, 40.2% statement previews, and 28.9% validations; selected programs comprised 55.8% base and 44.2% dependent DSL statements.Figure 12 visualizes the call and statement distributions across the evaluation benchmark.Source: B EXTENDED RESULTS · p. 19, Figure 12 · p. 21
  • Figure 11 extends the qualitative comparisons, while Figures 13–16 show the SUBSTANCE-50 and TOYS4K-50 evaluation assets.Figure 11's top row shows checkerboards and red seams; the bottom row shows parametrized UV atlases.Source: B EXTENDED RESULTS · p. 19, Figure 11 · p. 20, Figure 13 · p. 22
Loading 2610.02507v1…