MeshQuery uses a training-free VLM agent to plan UV seams, cutting charts by 2.9x/4.29x and seam length by 1.63x/1.7x versus the strongest baseline on Substance 3D and Toys4K, with artists preferring its results in 80.9% of comparisons

It introduces a training-free agentic UV unwrapping pipeline in which a Vision-Language Model plans 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. Unlike routes that rely on training or autoregressive seam prediction, the method decouples high-level intent planning from low-level edge selection and runs without training. The abstract states the approach is training-free and describes the VLM planner, the edge-selection tool set, the natural-language domain knowledge, and the feedback loop.

It designs a queryable mesh representation together with a domain-specific language (DSL) that lets the agent 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. Seam planning moves from per-edge prediction to a compact program paired with an on-demand queryable mesh representation, making the plan expressible and iteratively refinable. The abstract directly states the queryable mesh representation, the DSL, the edge-selection operator program, and iterative refinement 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. It reports quantitative comparisons against the strongest baseline plus a professional-artist preference rate across two datasets. The abstract reports specific multipliers and an 80.9% artist preference rate, but does not give sample sizes, rating protocols, or statistical test details in the visible text.

Decoupling high-level intent planning from low-level edge selection and using a compact mesh representation lets MeshQuery run on different backend VLMs and scale to meshes an order of magnitude larger than autoregressive seam prediction. Backend substitutability and scale extensibility are presented as consequences of the design, pointing to a larger mesh range than autoregressive seam prediction. The abstract states this with 'Ultimately', as the authors' summary of what the design decoupling enables.

The work targets automatic UV unwrapping of production-grade quad meshes, suited to graphics content production settings that require seam planning and UV layout. Its setting has a VLM plan under natural-language UV-unwrapping domain knowledge, using edge-selection tools and a queryable mesh representation, and iteratively refine from UV quality feedback. The evaluation described in the abstract covers Adobe Substance 3D and Toys4K meshes, comparing chart counts and seam lengths against the strongest baseline and using professional artist preference as an evaluation dimension. The method is described as training-free, able to run on different backend VLMs, and able to scale to meshes an order of magnitude larger than autoregressive seam prediction, which opens room to plug the pipeline into different model backends and larger assets.

The visible text is abstract-level information and does not provide sample sizes, artist rating protocols, statistical tests, failure cases, or concrete comparisons across backend VLMs, so the distributions and robustness behind the 2.9x/4.29x chart reduction, the 1.63x/1.7x seam-length reduction, and the 80.9% preference rate remain open questions. The expressive limits of the queryable mesh representation and DSL, the convergence behavior of the feedback loop, and actual performance on meshes an order of magnitude larger also need confirmation in the full text. In addition, the abstract does not state applicability to non-quad meshes or non-production-grade assets.