acceptodds
Under review as a conference paper at ICLR 2027

MeshQuery: Agentic Seam Planning for UV Parameterization

13 views

Abstract

We present MESHQUERY, a training-free agentic approach to automatic UV un- wrapping 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 in- formation 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 2.9×/4.29× fewer charts and 1.63×/1.7× 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 VLM and scale to meshes an order of magnitude larger than autoregressive seam prediction

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI Agents, UV Seam Layout

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wYAVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles automatic UV seam placement. MeshQuery is a training-free framework where a VLM queries a preprocessed relational mesh database in SQL and writes its seam plan as a DSL program over aggregate edge-selection operators, refined from UV metrics. The innovation is the agent interface rather than new geometry, keeping token cost near-constant with mesh size. It gives 3-4x fewer charts and 1.6x shorter seams than PartUV, preferred by annotators in ~80% of comparisons.

Critical strengths
  1. A scalable agent interface. The VLM never touches raw geometry: SQL queries plus aggregate edge operators keep tool calls and tokens near-constant (36k vs 230k for OBJ), enabling ~300k-face meshes.
  2. Clear and accessible writing. The pipeline figure, the DSL example, and the appendix implementation details make a fairly intricate system easy to follow and, in principle, reimplementable.
Critical weaknesses
  1. Limited practicality: the method needs clean quad meshes, yet production assets often have irregular connectivity from which no reliable edge loops can be extracted.
  2. Core contribution unablated: no experiment shows that the queryable representation and DSL, rather than the VLM's intrinsic capability, drive the gains on this task.
Questions for authors
  1. Connectivity requirement. How does the method's performance degrade when the input mesh has poor edge flow?
  2. Semantic resolution. Was K = 8 ever varied? How do seam quality and failure rate change on assets with many parts, such as characters, garments or multi-part props?
  3. Role of the visual channel. What do render_view, get_pick_view, pick_points and preview return, and when are they invoked? How much do the metrics change if all image inputs are removed?
  4. FlattenAnything configuration. Figure 4 shows roughly 14k charts, which contradicts that method's low-chart design. Are the reported sampling and iteration counts its official defaults at this scale, and can you reproduce its published behavior on its own evaluation assets before transferring the configuration here?
5 comments
AuthorsOpenReview ↗

Answers to wYAV

W1, Q1. Limited practicality / Connectivity requirement.
While our method leverages quad meshes (the default in industry DCC pipelines), we show that it extends to irregular triangle meshes (Fig. 14) by replacing topological loop operators with SeamCut (Lucquin et al. 2017) at the cost of additional time. Across five triangle meshes, on average, 44.40% of interior vertices deviate from the regular valence of 6, while 9.62% of faces have minimum angles below 10◦, indicating irregular connectivity and poor triangle quality.
We kindly refer to our General Comment for conclusions and action items related to this comment.

W2. Core contribution ablated.
We ablate the core contribution by providing the VLM backbone with the prompt “We want to UV-unwrap this mesh most professionally. List indices of the edges that we should mark as seams.”, the OBJ path and an offline shell with Blender (no queryable representation, labels, DSL, MCP tools, images, evaluator feedback or gates). We report the three Toys4K-50 quad meshes subset. We see that frontier Opus 5 can provide a seam plan for the task without our framework. However, it is less efficient, slower, and more expensive. On the opposite, Qwen3.8-27 B cannot solve the task without our framework: no final answer contains an edge list. We conclude that the proposed framework enables the frontier Opus 5 backbone at a reduced cost, and enables smaller non-frontier Qwen backbone.
These findings are further extended with other backbones and ablations, merging multiple comments, as discussed in our General Comment. We therefore include conclusions and action items in the General Comment.

Q2. PartField K Variation for Text Semantic Resolution.
We explore varying PartField K ∈{4,8,16} tested on opus 5 in Table 13 over three random Toys4K- 50 subset meshes. The framework is robust to different values of K, and all metrics remain in a comparable range. We conclude that frontier backbones are more robust to text-prior quality and benefit in speed. From the core ablation in the general comment, smaller VLM backbones gain distortion performance from the textual prior.
□ We will include this ablation analysis in the appendix.

Table 13: PartField K ablation over three quad meshes.

KCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
45.36.55.10.4450.0000.6940.9170.400.1181.0861.068
86.06.95.80.4600.0000.7210.9190.470.1391.1021.080
165.77.05.30.4420.0000.7550.9450.380.2131.1631.058

Q3. Role of the visual channel.
In Table 12, we ablate the visual channel on three random toyz4k-50 meshes. Opus 5 remained effective, indicating strong abstraction of our mesh representation, while the smaller model Qwen3.5- 9B benefits more from visual support. We attribute this to Opus’s stronger reasoning abilities. In addition, we remark (Fig. 17, Appendix) that get_pick_view (21) and pick_points (7) are used much less than query (942) among the paper’s 100 benchmark runs. get_pick_view returns a PNG of the mesh from a given view, and pick_points annotates points projected into the figure for a geodesic seam.
□ We will include this ablation in the appendix.

Table 12: Opus 5 and Qwen3.5-9B visual feedback ablation.

AgentVariantCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
Opus 5Full pipeline6.06.95.80.4600.0000.7210.9190.470.1391.1021.080
No visual channel7.79.07.00.4560.0000.6830.9030.440.1201.0911.068
Qwen3.5-9BFull pipeline4.76.46.40.4220.0420.6070.8660.280.2721.2691.313
No visual channel3.06.54.40.4870.3670.5300.7650.260.4291.4421.609
AuthorsOpenReview ↗

Answers to wYAV

Q4. FlattenAnything configuration.
FAM run with the original configuration of 10,000 iterations, and we re-ran it with the recommended clean mesh() preprocessing (which we previously avoided as it alters the input mesh). We reproduce the test script with per-vertex UVs agreement of 5·10−9. In Table 8, we also reproduced evaluations on libigl library meshes (camel-head, cow, fandisk, fertility) using published values. Substance meshes may contain multiple sub-meshes, justifying FlattenAnything chart count higher than 1.0. In Table 9 we report our updated values for FlattenAnything on our dataset.
□ We will update Table 1 of the main paper with these values.

Table 8: Conformal distortion of FlattenAnything on reproducing four models of its paper. We use public libigl versions of the meshes.

ModelPaper reportedOur rerun
camel-head0.0880.147
cow0.2100.329
fandisk0.1050.124
fertility0.1660.222

Table 9: FlattenAnything on TOYS4K-50 and SUBSTANCE-50 10,000 iterations, with additional preprocesssing clean_mesh() (n) denotes the number of successful runs within the 30-minute time budget. FlattenAnything produces 1 UV island per mesh; therefore, seam quality metrics are unavailable. Substance meshes may contain multiple sub-meshes, justifying an average chart count higher than 1.0.

Toys4K-50

Input (n)UV charts ↓UV seam ↓3D seam ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
Published (49)13974.4108.1159.20.4990.0020.5090.9840.910.7611.59617.187
Welded rerun (50)1.0———0.940———1.4194.17711.709

Substance-50

Input (n)UV charts ↓UV seam ↓3D seam ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
Published (50)19.717.05.90.5450.2983.0420.9020.581.7034.06743.716
Welded rerun (37)2.4———1.000———1.6974.50158.199

We kindly let the reviewer know we have uploaded the revised PDF, incorporating the additional experiments and clarifications discussed in our responses. We would be grateful for the reviewer's view on whether our response to the weaknesses addresses the concerns. We recognize reviewers are busy, and we appreciate the time already spent on our submission.

Reviewer_NHd7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MeshQuery, a training-free agentic framework for automatic UV unwrapping of production-grade quad meshes, designed to generate artist-aligned seams by leveraging the reasoning capabilities of a Vision-Language Model (VLM). The authors introduce a queryable mesh representation and a Domain-Specific Language (DSL) that enable a VLM agent to perform high-level seam planning without directly processing individual mesh vertices and edges. First, the input mesh is converted into a relational database containing topological, geometric, and semantic attributes, where semantic labels are obtained using PartField segmentation and VLM-based labeling. The agent selectively retrieves task-relevant information through SQL queries, reducing the token overhead associated with high-resolution meshes. Second, the authors introduce a seam-selection DSL that expresses seam plans as compact programs composed of geometric and topological operators, such as edge loops, semantic boundaries, and geodesic paths. Rather than predicting seams directly, the VLM generates and previews these programs, while deterministic operators perform the underlying edge selection. Third, the authors incorporate a closed-loop feedback mechanism in which the generated seams are used for UV parameterization, and quality metrics such as distortion, seam length, and UV chart count guide iterative refinement. A seam-preference skill provides domain-specific artistic knowledge to guide the agent's planning strategy. Quantitative and qualitative experiments are conducted on Substance-50 and Toys4K-50, against optimization-based and learning-based baselines. MeshQuery achieves 2.90×/4.29× fewer UV charts and 1.63×/1.69× shorter UV seams than PartUV on the respective datasets, while 10 professional artists prefer its results in 80.9% of A/B study against PartUV. The authors also provide ablation studies demonstrating the importance of the feedback loop and examining the influence of different VLM backbones.

Critical strengths

1. Well-motivated agentic formulation.

The paper introduces an interesting approach to UV seam planning by separating high-level semantic reasoning from low-level geometric operations. Rather than predicting individual seam edges, the VLM generates executable programs that leverage deterministic geometric operators. This formulation effectively combines the semantic understanding of VLMs with the precision of traditional geometry processing algorithms. MeshQuery does not require task-specific model training or a large dataset of artist-authored seams except off-the-shelf model (pretrained PartField). Its modular design, combining queryable representations, geometric tools, and natural-language skills, provides a pipeline that is likely to remain effective even as AI agents continue to evolve and are replaced by more advanced models.

2. Scalable and efficient mesh representation.

The proposed queryable mesh representation and DSL provide an effective interface between VLMs and high-resolution 3D meshes. By retrieving only task-relevant geometric, topological, and semantic information, the framework avoids directly serializing large meshes into the VLM context. The experiments demonstrate nearly constant observation-token consumption across different mesh resolutions and successful processing of meshes containing over 300K polygons. The method achieves substantial improvements over PartUV, reducing UV chart counts by 2.90×/4.29× and UV seam lengths by 1.63×/1.69× on Substance-50 and Toys4K-50, respectively. The resulting seam layouts are also visually coherent and often closely resemble artist-authored UV seams.

Critical weaknesses

If all of the following are resolved, I would be willing to increase my score.

1. High computational cost and dependence on powerful VLMs.

Despite its favorable scaling in observation-token consumption, the framework requires an average of 11.8 minutes for the agentic process, with additional preprocessing taking a median of 2.6 minutes. Approximately 71% of the agentic runtime is spent on VLM reasoning. The primary experiments rely on Claude Opus 5 with a large context window, while the smaller Qwen model exhibits noticeably worse distortion metrics. These results raise concerns about computational efficiency, practical deployment costs, and dependence on proprietary frontier models. Other agent engineering papers [1, 2, 3] experiment with various backbones and demonstrate the superiority of methodologies that are not dependent on the backbone. To demonstrate generalizability of proposed method, I recommend adding at least various versions of Claude models, different versions or sizes of Qwen, or other proprietary models like GPT-variants. (Add at least 2 of these models)

2. Ambiguous visual presentation.

The qualitative results of the current paper present only a textureless example. For topics where visual quality checks are important, there should be more samples showing how the actual texture corresponds to the UV map. For PartUV, the UV map results are color-mapped to a chart for about 30 unique meshes and displayed alongside the rendered mesh. On the other hand, the figures in this paper lack this level of detail, making it difficult to tell which part of the 3D model each chart corresponds to when the actual texture is applied. It would be good to add chart-based color mapping to each figure and include qualitative results for meshes with textures as well.

3. Limited generalization to arbitrary mesh topology.

The framework is primarily designed for quad meshes, where topological edge loops and other aggregate edge-selection operators are particularly effective. However, many real-world 3D assets, including generated 3D meshes, consist predominantly of triangles or irregular topology. Although the authors discuss possible extensions, the effectiveness of the proposed framework on general triangle meshes remains unverified. I wonder if it would have been better to mix quad meshes and triangle meshes when constructing the evaluation dataset. In the rebuttal, I would like to see qualitative examples for at least 4 triangle meshes.

4. Effect of segmentation is not studied

PartField may fail to produce accurate part segmentation for production meshes with complex topology, irregular connectivity, or challenging semantic structures. The performance of the framework under inaccurate semantic segmentation is also not systematically investigated, despite its reliance on PartField and VLM-generated semantic attributes. For 1 or 2 representative meshes with a suitable topology, I would like to see the start visualization from the initial UV map after performing PartField, and include visualization data showing how the UV map changes through a feedback loop.

[1] Cho et al., SpatialClaw: Rethinking the action interface for agentic spatial reasoning. NeurIPS 2026

[2] Chen et al., SpaceTools: Tool-Augmented Spatial Reasoning via Double Interactive RL, CVPR 2026

[3] Luo et al., pySpatial: Generating 3D Visual Programs for Zero-Shot Spatial Reasoning, ICLR 2026

Questions for authors
  1. It appears that only the statistics of the dataset used for evaluation in Figure 5 have been plotted. This can also be a useful statistic, but it would be better to analyze the tool calls, runtime, and token consumption of the methodology after subdividing or simplifying a mesh to adjust the number of vertices and faces. Since these meshes have the same underlying surface topology, I expect them to have the same shape of uv maps, so it would be good to check whether that is the case. If it is possible to simplify without compromising the topology, it is expected that preprocessing can reduce the total processing time.
8 comments
AuthorsOpenReview ↗

Answers to NHd7

The figures are available in the main PDF appendix.

W1. High computational cost and dependence on powerful VLMs.
We included evaluations over different backbone size variants (Claude Opus5.5, Claude-Sonnet-5, Qwen-9b).
We include the results, conclusion, and action items in our General Comment.

W2. Ambiguous visual presentation.
Following PartUV visualizations, we can report consistent UV chart colors to show correspondence on the 3D surface as exemplified in Fig 12).
□ We propose to add a new sibling figure to Figure 16, with this exact layout for the same example meshes.

W3. Limited generalization to arbitrary mesh topology.
In addition to the quantitative table in the general comments, we included qualitative Fig. 14. For triangle meshes, the pipeline relies on seam cuts as a substitute for topological loops, which are not applicable. Although wall clock time and jaggedness may increase due to the inherent characteristics of triangle meshes and our definition of seams as mesh edges, the pipeline can still effectively produce seams.
□ We propose to add this figure to the appendix, and kindly to our General Comment for other conclusions and action items related to this comment.

W4. Effect of segmentation is not studied.
Semantic labels provide only a textual prior for SQL query generation and edge-ID search, not a starting point initialization for UV seams. Fig 13 shows how an hypothetical PartField-boundary unwrapping differs from our pipeline. Uninformative labels (e.g., foot in the bowl sample) do not penalize geometric or topological queries and feedback-driven seam refinement.
□ We will add a clarifying remark in Sec.3.1 ”Mesh Attributes” to state that PartField is used only as semantic text prior to enriching the mesh representation, and that PartField-implied seams are not used as initialization.

Q1. Subdivided meshes.
We ablate the mesh resolution by upscaling (x2) via subdivision. Tab 2 shows that increasing mesh resolution from ×1 to ×2 (4× more faces, from 15K to 60K) preserves seam quality. In addition, the agent calls increase by only 9.5% (21 → 23), while active runtime increases by 2.29× (3.4 → 7.8 min), substantially less than the 4× increase in face count. Output UV atlases in Fig. 10 (appendix) appear visually similar.
□ We will extend the mesh-scaling analysis in Sec. 4 of the main paper with these results, and add a Figure 11 to the appendix.

Table 2: Opus 5 at ×1 (original) and ×2 (subdivided) over three Toys4K-50 meshes.

Res.FacesCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓CallsActive minOutput tok. (k)
×114,9576.06.915.850.4600.0000.7210.9190.470.1391.1021.080213.411.4
×259,8295.76.935.260.4490.0000.7280.9320.490.1231.0891.071237.817.9

We kindly let the reviewer know we have uploaded the revised PDF, incorporating the additional experiments and clarifications discussed in our responses. We would be grateful for the reviewer's view on whether our response to the weaknesses addresses the concerns. We recognize reviewers are busy, and we appreciate the time already spent on our submission.

Thanks for the authors’ response. I have read it as well as the other reviews.

It seems that the figure numbers are not matching what the authors are stating. Also, the figures are of degraded quality, making it difficult to see details. Could you please check again?

Reviewer_7DTAOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

MeshQuery is a training-free agent for UV seam planning on quad meshes. It reads precomputed mesh attributes and writes a seam plan in a small DSL of edge-selection operators, then revises the plan from UV-quality feedback until the quality gates pass or the 40-minute budget runs out. On two 50-asset sets it reports 2.9x/4.3x fewer charts and 1.6x/1.7x shorter seams than PartUV. Mean chart count is close to the artist reference on Substance (19.6 vs 19.5), and 10 professional artists prefer it in 80.9% of decisive ratings. Its text observations also grow slowly with mesh size, while raw OBJ grows almost linearly.

Critical strengths
  1. The layouts look artist-like. Mean chart count matches the artist reference on Substance (19.6 vs 19.5), and professionals prefer the output in 80.9% of decisive ratings against PartUV. The two are complementary, though a matching count says nothing about where seams go.
  2. Planning stays separate from mesh-sized edge selection. Observation volume grows slowly with mesh size, up to roughly 465x smaller than OBJ on the largest meshes, and the tested set reaches 330k vertices.
Critical weaknesses
  1. The trade-off is not measured on a common definition. The paper's per-face area gate, PartUV's chart-level threshold and OptCuts' Diriclet bound measure different things, so the numbers cannot be lined up. Table 1 also gives the method the highest area distortion among ranked methods, and OptCuts' 3.8 charts on 38 successful cases sit next to 19.6 on all 50. Excluding failures is fair, but there is no matched subset aggregate.
  2. Several reporting gaps block a precise read. Area distortion is 1.178 in Table 1 and 1.151 in Table 2 for the same setting; the statement split is 55.8/44.2 in one place and 60.0/40.0 in another; and 342 plan versions with a median of 3 sit badly with 66% first-attempt acceptance. The study also gives no tie counts, no blinding, and no uncertainty over repeated ratings.
Questions for authors
  1. Which statistic really decides acceptance, including normalization and aggregation? Which OptCuts mode and bijectivity setting did you use? If existing outputs allow it, could you show the trade-off on matched successful cases while keeping failure rates visible?
  2. Could you reconcile the two distortion values and the two statement splits, and say whether they come from the same runs? How do plan versions map onto the stopping rule, given that 66% first-attempt acceptance would imply a median of one version?
  3. Could you report wins, losses and ties by dataset and rater group, and say whether method identity was hidden and order randomized? A mesh-level summary would be more convincing than the rating-level figure.
3 comments
AuthorsOpenReview ↗

Answers to 7DTA

Q1. Acceptance statistic.
A plan is accepted when its UV atlas passes the distortion threshold, and no face has zero UV area (no collapsed faces are allowed) PartUV’s chart-level threshold and OptCuts’ Dirichlet bound are internal constraints of those methods and are not used for acceptance.
□ We will add this clarification in section 3.

Q1. OptCuts configuration.
We ran the full OptCuts method (method type 0), not the EBCuts, no-dual, or distortion-only variants, with bijectivity enabled and authors’ defaults: initial λ= 0.999, symmetric Dirichlet distortion bound bd = 4.1, the default initial cut (OptCuts bin 100 <mesh> 0.999 1 0 4.1 1 0).
□ We propose adding these details to the main paper, Sec. 4.1, ”Implementation details.”

Q1,W1, The trade-off is not measured on a common definition. In Tab 11, we provide the aggregate over the meshes where every method returns a solution: 37 Toys4K meshes, and 27 SUBSTANCE meshes. The retained meshes have lower median vertex counts than the excluded meshes in each dataset: Substance 5,474 (27 retained) vs. 16,113 (23 excluded); Toys4K 3,984 (37) vs. 111,618 (13). Substance meshes may contain multiple sub-meshes, justifying FlattenAnything chart count higher than 1.0.
□ We propose to add these results in the Appendix.
We also report in Table 10 PartUV at different distortion thresholds besides the recommedned default τ=1.25 to τ=1.5 and τ=2.0. We observe that our method yields higher seam placement quality than PartUV under all conditions, possibly because it is not conditioned on artist seam-placement priors.
□ We propose to mention this in the main paper, backed by Table 10, in Appendix.

Table 11: Quantitative seam quality metrics: bold is best and underline the second-best values. FlattenAnything and the artist reference are not ranked.

Substance: 27 matched meshes

MethodCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
FlattenAnything+cleanmesh2.4———1.000———1.6914.41862.077
OptCuts2.86.33.20.5000.0670.4500.8110.910.1971.1841.134
xatlas (LSCM)175.734.716.60.4950.0630.4290.8551.040.1761.1031.114
Blender Smart UV108.132.915.90.5190.0800.4200.8101.010.1601.1011.164
PartUV41.619.39.80.4870.0680.4540.7890.590.1521.1101.033
Ours14.312.05.30.5140.0220.6100.8900.690.1701.1421.082
Artist reference17.113.46.20.5460.0840.5220.8770.450.1441.0971.071

Toyz4K: 37 matched meshes

MethodCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
FlattenAnything+cleanmesh1.0———0.973———1.3823.72810.100
OptCuts1.54.63.70.4900.1980.4130.7860.530.1901.1831.140
xatlas (LSCM)49.121.014.40.5250.0410.4050.8240.670.1241.0881.295
Blender Smart UV39.520.514.20.5260.0640.3980.8120.680.1551.0981.151
PartUV11.011.47.90.5110.1040.4310.8000.660.1941.1431.059
Ours8.78.75.80.5460.0670.6250.9080.450.2151.1981.188

Table 10: Toys4K-50 comparison between different PartUV thresholds. Bold denotes the best and underline the second-best values.

MethodCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓
Ours (Opus 5)9.38.96.10.5330.0630.6100.9040.470.2151.2011.172
PartUV (default, τ=1.25)39.915.010.10.5010.1080.4250.8040.820.1921.1431.054
PartUV (τ=1.5)21.012.58.40.5020.1480.4340.7950.720.2631.2191.059
PartUV (τ=2.0)12.211.37.60.4930.1500.4370.7850.700.3411.2991.072
AuthorsOpenReview ↗

Answers to 7DTA

Q2, W2a, Reporting gaps: area distortion.
Yes, same runs and plans. The gap is due to one mesh, hammerhead shark, because of a mistake in our export for 28 submeshes representing its teeth. After fixing the export, the correct Substance area distortion is 1.160.
□ We will fix these values in both tables.

Q2, W2b, Reporting gaps.
Both come from the same runs. One figure reports all 1,911 statements the agent submitted: 58.4% Base / 41.6% Dependent; the other reports 1,174 statements in the 100 selected programs: 60.0% / 40.0%.
□ We will update the main paper with a 58.4% / 41.6% split.

Q2, Reporting gaps: plan versions.
The 66% counts only the distortion check on the first plan version. Once the quality criteria hold, the agents attempt a secondary refinement step to improve island convexity. Ultimately, the plan is selected via lexicographical ranking.
□ We will add this clarification in Sec. 3., as well as harmonize Figure 17 and sibling appendix B ”Tool usage statistics”.

Q3 Reporting gaps: user study.
We show in Figure 15 an example for the human study. It is a blind A–B test with a hidden method name, and both candidates of a mesh were rendered with identical checker texture, lighting, and cameras. Each rater received a balanced subset of 40 out of 40 of the 80 meshes (20 per dataset) in a random order, so all 15 sessions had distinct subsets and orders. Each mesh was rated by 3–12 raters. Our method is preferred in 79.8% of decisive ratings (95% CI 70.0–88.3); for the 10 artists, 80.9% (69.7–90.4). The mean per-mesh preference for our method is 76.9% (mesh-bootstrap 95% CI 69.5–83.6). The result holds on both datasets (Substance 29/8/3, Toys4K 31/6/3) and for artists alone (61/14/5). Agreement between raters on the same mesh is moderate (K α= 0.38; 69.1% of rater pairs give the same answer).
□ We propose to specify in Sec.4.2 ”Human evaluation results” that the methods are hidden and order randomized, plus add this detailed table and the study screenshot in the Appendix.

Table 14: Human study, our method vs. PartUV. Meshes O/P/S: meshes whose majority of decisive ratings prefer ours / PartUV / split.

RatersDatasetnWinsLossesTiesWin% [95% CI]Meshes O/P/S
All raters (15)Overall60039610010479.8 [70.0, 88.3]60/14/6
Substance300184575976.3 [63.1, 87.5]29/8/3
Toys4K300212434583.1 [69.8, 93.6]31/6/3
Artists (10)Overall400271646580.9 [69.7, 90.4]61/14/5
Substance200127393476.5 [61.2, 89.2]29/7/4
Toys4K200144253185.2 [70.8, 96.3]32/7/1
Non-artists (5)Overall200125363977.6 [65.9, 87.7]57/13/9
Substance10057182576.0 [58.8, 90.1]27/7/5
Toys4K10068181479.1 [62.9, 92.8]30/6/4

We kindly let the reviewer know we have uploaded the revised PDF, incorporating the additional experiments and clarifications discussed in our responses. We would be grateful for the reviewer's view on whether our response to the weaknesses addresses the concerns. We recognize reviewers are busy, and we appreciate the time already spent on our submission.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Comment

We thank the reviewers for their constructive feedback. “The layouts look artist-like” (7DTA) and “often closely resemble artist-authored UV seams” (NHd7). “This formulation effectively combines the semantic understanding of VLMs with the precision of traditional geometry processing algorithms” (NHd7). “The proposed queryable mesh representation and DSL provide an effective interface between VLMs and high-resolution 3D meshes” (NHd7), “a scalable agent interface” (wYAV).
In response, we extended the results with additional backbones, triangle meshes with irregular connectivity, and additional ablations.

All tables and figures are also attached in blue into our PDF.

In summary:

  1. Core ablations and beyond powerful VLMs: our framework reduces costs for frontier VLMs, while enabling small VLMs. Since submission, we ran additional comparisons. We compare models Opus 5 (paper), Sonnet-5, Qwen3.8-27B (paper), Qwen3.5-9B, backbones on three random quad meshes from Toys4K-50, across three settings: bare prompt + .obj reference, our framework without semantic text priors, our full framework (including semantic text priors). Observations are twofold: on the one hand, we are able to cut costs by up to 51% and clock time by up to 69% and improve distortion (area by 11%, texel spread by 28%) on frontier Opus 5 using our framework, even if the baseline can solve the task too. On the other hand, our framework pushes lower-end models from failure to success, or significantly reduces distortion. Using semantic text priors is beneficial, especially in terms of time and cost for frontier models. See Tab. 4.
    □ We propose enriching our original supplemental material with these figures and adding a dedicated paragraph in Section 4.2 of the main body to discuss this generalization.

  2. Triangle meshes are supported. However, our framework best exploits quad-dominant meshes, as typically found in production (e.g., most assets in the Blender asset library (https://www.blender.org/download/demo-files), used in short film productions, are quad meshes). To demonstrate this, we additionally ran Opus 5 (paper), Opus 5.5, Qwen3.8-27B (paper), and Kimi K3 backbone models on five random triangle meshes from Toyz4K-50. In this setup, the framework relies on SeamCut instead of topological loops, which are not applicable in the triangle-mesh setting. We observe that all backbones succeed on all meshes, with seam quality in the same range as for quad meshes. The metrics most negatively affected are clock time and jaggedness, consistent with the increased difficulty of finding clean loops. See Tab. 3.
    □ We propose enriching our original supplemental material with these figures and adding a dedicated paragraph in Section 4.2 of the main body to mention this generalization.

  3. We also address the remaining questions, with more details in the dedicated per-review threads.

  • ambiguous visual presentations: □ we propose to add visualizations using with per- island color, similar to PartUV NHd7 (W2).
  • PartField segmentation usage clarification: □ we propose to add a clarifying remark in section 4.1 ”implementation details” to state that PartField is used only as semantic text prior to enriching the mesh representation, and remark that PartField-boundaries are not used as seam initialization. NHd7 (W4).
  • mesh downscaling ablation: □ we propose to complement Figure 5 and paragraph ”Mesh size performance scaling.” with results from the experiment described in the comment on reviewer NHd7(Q1).
  • we outperform PartUV at matched distortion thresholds: □ we propose to mention this in Appendix A.4 and back up with figures reported in the comment on reviewer 7DTA(Q2).
  • visual channel ablation: smaller VLM backbones benefit more from visual feedback than frontier models. □ we propose to mention this in the Appendix B. and back up with figures reported in the comment on reviewer wYAV(Q3).

Core framework ablation. We compare different VLM backbones operating directly on an OBJ file with an offline Blender shell (baseline) against our framework, with and without text priors (semantic part labels), on 10 randomly selected quad meshes from TOYS4K-50. Percentages indicate changes relative to their baseline (green: better, not reported when baseline is not valid).

BackboneConfigurationTexel spread ↓Area dist. ↓Conf. dist. ↓Wall s ↓Output tokens ↓USD ↓
Opus 5Baseline (OBJ + Blender)0.3061.3081.1531,04974,0384.80
+ Pipeline (no text priors)0.201 (−34%)1.165 (−11%)1.113 (−3.5%)418 (−60%)23,807 (−68%)2.49 (−48%)
+ Pipeline (text priors)0.222 (−28%)1.201 (−8.2%)1.119 (−2.9%)385 (−63%)22,593 (−69%)2.34 (−51%)
Sonnet 5Baseline (OBJ + Blender)2.8414.059>100047737,9110.82
+ Pipeline (no text priors)0.228 (−92%)1.192 (−71%)1.183 (−100%)58740,4911.42
+ Pipeline (text priors)0.228 (−92%)1.184 (−71%)1.222 (−100%)44730,7131.31
Qwen3.8-27BBaseline (OBJ + Blender)FailsFailsFails1,23820,079—
+ Pipeline (no text priors)0.2901.3131.2661,08516,095—
+ Pipeline (text priors)0.2631.2441.24777411,061—
Qwen3.5-9BBaseline (OBJ + Blender)4.0029.050>100037511,715—
+ Pipeline (no text priors)0.457 (−89%)1.676 (−81%)1.931 (−100%)1,00125,437—
+ Pipeline (text priors)0.605 (−85%)1.888 (−79%)1.974 (−100%)1,05520,179—

Triangle mesh evaluation. We compare various VLM backbones on five triangle Toys4K meshes. Seam quality is in the same range as quads, except for clock time and jaggedness.

BackboneCharts ↓UV seam length ↓3D seam length ↓AO ↑Non-disk ↓Compact. ↑Convex. ↑Jagged. ↓Texel spread ↓Area dist. ↓Conf. dist. ↓Wall min ↓
Opus 511.610.05.00.5370.0000.5740.8930.890.2191.2641.16516.6
Opus 5.514.010.25.20.5300.0000.6070.9030.910.1881.2171.1359.7
Qwen3.8-27B12.211.85.60.4850.0000.5290.8721.180.2711.4281.24917.1
Kimi K314.810.15.60.5240.0050.5330.8870.950.1901.2281.13116.4

We kindly let the reviewer know we have uploaded the revised PDF, incorporating the additional experiments and clarifications discussed in our responses. We would be grateful for the reviewer's view on whether our response to the weaknesses addresses the concerns. We recognize reviewers are busy, and we appreciate the time already spent on our submission.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.