MeshQuery: Agentic Seam Planning for UV Parametrization

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

cs.CVOct 1, 2026

MeshQuery: Agentic Seam Planning for UV Parametrization

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

Organizations: Adobe Research, France

Abstract

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

Figures & tables

Figure 1: MeshQuery generates artist-aligned UV seams. Conversely to hand-crafted geometric optimization, MeshQuery enables a VLM agent to plan seams the way an artist does: it iteratively refines seams expressed as programs in a Domain-Specific Language (DSL), based on unwrapping feedback and on the mesh’s geometric, topological, semantic attributes, and downstream needs.

Figure 2: Agentic UV seam planning. The input mesh $M$ is first converted into a queryable representation $M^{*}$ that stores geometric, topological, and semantic attributes. Then a VLM agent, conditioned on seam-preference knowledge, iterates a three-stage loop: it observes $M^{*}$ through SQL queries and seam-selection statement previews, then generates a cohesive program $P$ , evaluates the plan through its UV quality metrics $q$ , and iteratively refines it until all quality bounds are met.

Figure 3: A domain-specific language (DSL) allows expressing edge selection and plan manipulation via a high-level seam selection program (JSON) based on aggregate edge selection operators.

Figure 4: Qualitative comparison on Toys4K-50 . For each asset, the top row shows a UV checkerboard and generated seams in red, while the bottom row shows the parametrized UV atlas.

Substance-50
	Seam quality								Distortion
Method ( $n$ )	UV Charts $\downarrow$	UV Seam length $\downarrow$	3D Seam length $\downarrow$	AO $\uparrow$	Non-disk $\downarrow$	Compact. $\uparrow$	Convex. $\uparrow$	Jagged. $\downarrow$	Texel spread $\downarrow$	Area dist. $\downarrow$	Conformal dist. $\downarrow$
FlattenAnything Zhang et al. (2024) ‡ (50)	(19.7)	(17.0)	(5.9)	(0.545)	(0.298)	(3.042)	(0.902)	(0.58)	(1.703)	(4.067)	(43.716)
OptCuts Li et al. (2018) † (38)	(3.8)	(6.3)	(3.4)	(0.504)	(0.051)	(0.473)	(0.819)	(0.84)	(0.197)	(1.172)	(1.133)
xatlas (LSCM) Lévy et al. (2002) (50)	355.5	44.2	26.6	0.503	0.065	0.428	0.853	1.21	0.163	1.095	1.111
Blender Smart UV Sheffer et al. (2007) (50)	162.6	37.9	21.5	0.521	0.071	0.408	0.807	1.14	0.159	1.102	1.162

Table 1: UV parameterization benchmark . Here,

(n)

is the number of successful runs. We highlight in bold and underlined values that indicate the best and second-best results, respectively. Methods producing failure cases ( † ) or excessively high mean distortion ( ‡ ) are excluded from the ranking.

Figure 5: Tool calls, runtime, and token consumption across varying mesh polygon sizes. Left : Colored paths connect meshes from the same semantic class, but with different geometry and topology. Tool call exhibits high-level agent reasoning, has negligible correlation with mesh size, while runtime scales super-linearly due to parameterization feedback cost per iteration. Right : VLM token cost across varying mesh polygon sizes is near-constant, while direct mesh serialization grows linearly.

Variant		Area dist. $\downarrow$	Angular dist. $\uparrow$	Conformal dist. $\downarrow$	Texel spread $\downarrow$
(a) Closed-loop feedback
Substance-50	With feedback (ref.)	1.151	0.928	1.100	0.178
	No closed loop	1.341 $+16.5\%$	0.887 $-4.4\%$	1.419 $+29.0\%$	0.377 $+111.6\%$
Toys4K-50	With feedback (ref.)	1.201	0.903	1.173	0.215
	No closed loop	1.525 $+27.0\%$	0.822 $-8.9\%$	1.359 $+15.9\%$	0.512 $+137.8\%$
(b) LLM model

Table 2: Ablations. We ablate the feedback loop (a) and the LLM model (b) and report distortion metrics (red indicates regression).

Figure 6: Closed-loop feedback. (Left) The cone of the rock chart fails distortion bounds. After refinement, the agent adds one apex-to-base slit, opening the cone into a fan while leaving the base chart untouched. Right : a penguin model with the generated seams at each iteration.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Property	Substance-50	Toys4K-50
Vertices	1,344 / 9,233 / 53,638	62 / 4,751 / 330,654
Object parts	1 / 1 / 6	1 / 1 / 1
Connected components	1 / 2 / 31	1 / 1 / 1
Triangle fraction	0 / 0.003 / 0.095	0 / 0 / 0.050

Appendix

Table 3: Dataset statistics. Entries are minimum / median / maximum over their respective set.

Figure 7: Dataset mesh size comparison. We show the overall distribution of mesh sizes across common 3D mesh datasets Jacobson and contributors (2013) ; Koch et al. (2019) ; Deitke et al. (2023) ; Yang et al. (2024b) ; Xiang et al. (2025) . Substance-50 and Toys4k-50 have comparable mesh size distribution with respect to the others, measured by the number of vertices.

Tool	Details	Time
1. Perception
Query
query (SQL) $\rightarrow$ SQL-edge $\|$ SQL-face, Stats	Is the main tool of the agent for collecting mesh attribute observations. Returns valid SQL-edges or SQL-faces fitting the query, as well as their aggregate distribution statistics.	$O(n)$
DSL statements Base : evaluated from SQL selections
topological_loop (SQL-edge) $\rightarrow$ Base	from an input SQL-Edge, it proceeds through a topological walk over the mesh and returns the set of connected edges. the chain that goes straight through every valence-4 vertex via the opposite edge, ending at poles or the boundary	$O(n)$
boundary_loop (SQL-face) $\rightarrow$ Base	returns Mesh boundary loop as the edges with exactly one incident face, chained into cycles	$O(n)$

Appendix

Table 4: Agent Tool interface. “?” marks an optional argument; “

\rightarrow

” gives the return type. A Base statement is evaluated from SQL selections alone; a Dependent statement takes the output of a Base statement as an argument, so a program is a two-layer dependency graph with in-degree at most one. Time complexity is given in

n

, the face count of the mesh;

m

is the number of statements in the submitted program and

r

the number of iterations of the flattener.

Figure 8: Relational mesh representation ( mesh.db ). PK and FK denote primary and foreign keys. Solid arrows indicate one-to-many relations; the dashed arrow links a candidate to its seed edge. Geometric signals are stored per face and per edge.

Figure 9: VLM semantic labeling . (a) Input : PartField splits the mesh into $K=8$ clusters based on the mesh semantics and geometry. (b) Output : the VLM produces a final set of labels for each face. The Bottom row illustrates the labeling process. For each k-th cluster, we render the mesh with the part colored red. Then we prompt the VLM to inspect the image and provide a text label for the selected area. We set the VLM to coin new terms only when strictly necessary, or reuse previously used ones to compose a label vocabulary.

Figure 10: Distortion performance scaling over mesh size. MeshQuery distortion performance does not degrade over larger meshes.

Figure 11: Qualitative comparison on Substance-50 . For each asset, the top row shows a UV checkerboard and generated seams in red, while the bottom row shows the parametrized UV atlas.

Figure 12: Tool use across both evaluation sets (100 runs). Left: Distribution of the 3,049 tool calls by tool and call kind. Right: Distribution of the 1,174 DSL statements in the 100 selected programs, grouped by the type of edge selection returned.

Figure 13: Substance-50 evaluation assets.

Figure 14: Substance-50 evaluation assets (continued).

Figure 16: Toys4K-50 evaluation assets (continued).

Vision-Language Agents

Vision-language agents use visual and language models to perceive environments, reason about goals, and take actions through interactive interfaces.

See all 44 papers

DreamUV: Unwrap Artist-like UV by End-to-End Flow Matching
Jun 21, 2026Quanyuan Ruan, Jiabao Lei, Xingyi Du +1Generative Modeling Conditional Generative Modeling
South China University of Technology · School of Data Science, The Chinese University of Hong Kong, Shenzhen · Lightspeed
MeshTailor: Cutting Seams via Generative Mesh Traversal
Mar 28, 2026Xueqi Ma, Xingguang Yan, Congyue Zhang +1Pointer Networks 3D Mesh Generation
Shenzhen University · Simon Fraser University
SeamFlow: Structure-Aware Flow Matching on Edge Probabilities for Artist-Like UV Unwrapping
Sep 4, 2026Yuming Zhao, Zangyueyang Xian, Qijian Zhang +4Flow-Based Generative Modeling Flow Matching
Meshy AI, USA · Nanyang Technological University, Singapore

Jun 21, 2026J/K move · Enter open · S save

DreamUV: Unwrap Artist-like UV by End-to-End Flow Matching

Quanyuan Ruan, Jiabao Lei, Xingyi Du, Xifeng Gao

South China University of Technology · School of Data Science, The Chinese University of Hong Kong, Shenzhen · Lightspeed

Generative Modeling Conditional Generative Modeling

Open paper AskPDF arXiv

UV parameterization is a fundamental step in 3D content creation, yet producing production-ready UV layouts remains challenging due to the gap between geometric distortion objectives and the stylistic preferences of professional artists. While classical methods optimize handcrafted energy functions, artist-authored UVs exhibit structural patterns such as straightened seams, axis-aligned islands, and flexible interior deformation, properties that are difficult to explicitly formulate. In this work, we present DreamUV, an end-to-end learning framework that formulates UV unwrapping as a generative Flow Matching problem. Rather than predicting a single optimal parameterization, DreamUV learns a mesh-conditioned transport process that maps noise samples to a distribution of artist-like UV layouts. To reflect real-world authoring practices, we introduce a boundary-aware training strategy that prioritizes seam geometry, and a Model-in-the-Loop Finetuning(MITL) scheme that explicitly accounts for discretization errors during sampling and stabilizes transport dynamics under heterogeneous supervision. We evaluate DreamUV on a large-scale dataset of professionally authored UV layouts. Experiments demonstrate that our method produces significantly straighter boundaries and tighter axis-aligned islands than both classical and learning-based baselines, while maintaining competitive distortion metrics. Qualitative results and a user study with professional artists further confirm that DreamUV generates UV layouts that are not only valid, but aligned with practical production requirements.

Abstract

Vision-Language Agents

Explore similar work

DreamUV: Unwrap Artist-like UV by End-to-End Flow Matching

MeshTailor: Cutting Seams via Generative Mesh Traversal

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