ArtUV: Artist-Style Neural UV Parameterization

• ArtUV is a neural UV unwrapping approach that redefines UV mapping by emphasizing semantic seams, clean island geometry, and downstream editability.
• The system employs a two-stage pipeline with SeamGPT for surface seam prediction and a learned refinement network to transform rough UV mappings into artist-style layouts.
• Empirical evaluations show ArtUV achieving superior UV utilization and low distortion while effectively balancing reconstruction, silhouette, distortion, and overlap objectives.

ArtUV denotes a line of research in neural UV parameterization that treats UV unwrapping not merely as a distortion-minimization problem, but as a production-oriented layout problem in which semantic seams, clean island geometry, efficient atlas use, and downstream editability matter explicitly. In its most specific usage, ArtUV refers to the fully automated, end-to-end “artist-style” UV unwrapping system that divides the task into surface seam prediction and artist-style UV parameterization refinement (Chen et al., 25 Sep 2025). The name is also used for an unsupervised, differentiable framework that augments geometry-preserving UV learning with semantic-awareness and visibility-awareness objectives (Zamani et al., 29 Sep 2025). Both formulations occupy the same broader design space: automatic UV generation for professional rendering pipelines in which geometric validity is necessary but not sufficient.

1. Problem definition and scope

In the artist-style UV unwrapping formulation, the input is a 3D mesh $M$ with vertices $V\in\mathbb{R}^{N\times 3}$ and triangle faces $F\in\mathbb{R}^{M\times 3}$, and the task is to map each 3D vertex to 2D coordinates $(u,v)$ in a UV plane while preserving topology and avoiding overlaps (Chen et al., 25 Sep 2025). The central claim is that production-quality UVs are judged by criteria that extend beyond bijectivity and low distortion. The paper identifies higher-level requirements including clean boundaries, efficient space utilization, semantic coherence, and readiness for 2D editing, all of which directly affect texture painting, part reuse, and later editing stages (Chen et al., 25 Sep 2025).

This framing is explicitly critical of several prior families of methods. Top-down methods such as LSCM and ABF++ can produce good parameterizations when given reasonable seams, but they remain dependent on expert seam design and manual cleanup. Bottom-up optimizers such as OptCuts jointly decide seams and parameterization, but may yield fragmented or semantically arbitrary charts. Learning-based systems such as Nuvo and FAM are described as still suffering from weak semantic awareness, time-consuming training or optimization for each scene or model, or poor topology and overlap behavior in the resulting UV maps (Chen et al., 25 Sep 2025, Srinivasan et al., 2023).

A separate ArtUV formulation generalizes this critique. It argues that classical parameterization criteria such as conformality, area preservation, bijectivity, and low distortion ignore two perceptually important properties: semantic chart alignment and low-visibility seam placement (Zamani et al., 29 Sep 2025). Taken together, these works define ArtUV less as a single algorithmic motif than as a shift in objective: from geometry-only UV parameterization toward atlas structures that are more interpretable and operationally useful in content creation.

2. Two-stage artist-style pipeline

The two-stage ArtUV system is designed to imitate a professional workflow. The first stage is surface seam prediction, implemented with SeamGPT, and the second stage is artist-style UV parameterization, implemented as a learned refinement of a rough optimization-based UV initialization (Chen et al., 25 Sep 2025).

SeamGPT formulates surface cutting as a sequence prediction problem. Cut vertices are spatially sorted and quantized, and each coordinate token is predicted autoregressively. For an input mesh, point clouds are sampled from vertices and edges and encoded into a latent shape condition with a point cloud encoder, after which an hourglass-like autoregressive Transformer decoder predicts coordinate tokens from a start token to an end token. The predicted discrete coordinates are then projected back onto the mesh to obtain seam vertices. The stated purpose is to generate semantically meaningful cutting seams rather than arbitrary or purely local cuts (Chen et al., 25 Sep 2025).

The second stage does not attempt to generate UVs from scratch. Instead, a rough UV map $Q_i$, obtained in the implementation from Blender’s Ministretch, is combined with mesh information $I_M$ and passed to a learned refinement model. The paper writes the combined input as $I = I_M + Q_i$, where the mesh features include vertices, faces, normals, degrees, and curvature. The model predicts a UV offset $Q_o$, and the final output is given by

$Q_{\text{pred}} = Q_i + Q_o.$

This residual formulation is central to the method: the network learns how to improve an existing mapping into an artist-style UV layout rather than solving the full parameterization problem ab initio (Chen et al., 25 Sep 2025).

Architecturally, the refinement network first processes inputs with a residual MLP with adaptive dimensional mapping, propagates local structure through graph convolutions using SAGEConv over a mesh graph, and then uses a Pyramid Encoder-Decoder to capture global interactions with attention and coarse-to-fine decoding. The output is a 2D offset field added to the initial UVs. The paper emphasizes “semantic consistency” and “topological structure preservation” as the operational consequences of this design, especially in contrast to point-cloud-based formulations that do not preserve mesh adjacency (Chen et al., 25 Sep 2025).

3. Objective function and style regularization

The parameterization stage is supervised against artist-optimized target UVs $Q_{gt}$. Before reconstruction error is measured, the initial UV coordinates $V\in\mathbb{R}^{N\times 3}$0 are aligned to $V\in\mathbb{R}^{N\times 3}$1 using Horn’s closed-form absolute orientation method. The covariance matrix is

$V\in\mathbb{R}^{N\times 3}$2

followed by

$V\in\mathbb{R}^{N\times 3}$3

$V\in\mathbb{R}^{N\times 3}$4

$V\in\mathbb{R}^{N\times 3}$5

This alignment is used before computing losses so that refinement is evaluated against the intended artist-style layout rather than an arbitrary global rotation (Chen et al., 25 Sep 2025).

The main reconstruction term is

$V\in\mathbb{R}^{N\times 3}$6

To encourage tidy island outlines, ArtUV adds an edge-preserving silhouette term based on differentiably rendered silhouettes of the predicted and ground-truth UV maps,

$V\in\mathbb{R}^{N\times 3}$7

For distortion control, the method computes per-face Jacobians of the 3D-to-2D map and uses the singular values $V\in\mathbb{R}^{N\times 3}$8 and $V\in\mathbb{R}^{N\times 3}$9 as distortion indicators, with the stated goal of penalizing differences between them so that lower values correspond to more conformal, less stretched parameterizations. To prevent overlaps, it adds an overlap penalty based on flipped face normals in UV space,

$F\in\mathbb{R}^{M\times 3}$0

where $F\in\mathbb{R}^{M\times 3}$1 is the face normal in UV space and $F\in\mathbb{R}^{M\times 3}$2 is the reference viewing direction (Chen et al., 25 Sep 2025).

The total objective is

$F\in\mathbb{R}^{M\times 3}$3

with implementation weights $F\in\mathbb{R}^{M\times 3}$4, $F\in\mathbb{R}^{M\times 3}$5, $F\in\mathbb{R}^{M\times 3}$6, and $F\in\mathbb{R}^{M\times 3}$7 (Chen et al., 25 Sep 2025). The paper interprets this weighting as prioritizing reconstruction and silhouette shape as the main style targets while retaining explicit regularization on distortion and overlap.

A similarly named ArtUV framework uses a different loss construction. Its unsupervised backbone combines wrapping, cycle-consistency, anti-overlap, and differential and triangle-based distortion losses, then augments them either with semantic partitioning or an AO-weighted seam objective that minimizes the mean ambient occlusion of seam vertices (Zamani et al., 29 Sep 2025). The coexistence of these two designs clarifies that “artist-style” in this literature can be approached either through supervised refinement toward artist-authored targets or through unsupervised geometric learning augmented with perceptual objectives.

4. Data construction and evaluation protocol

A major part of the supervised ArtUV system is the ArtUV-200K dataset. The authors collect UV-mapped meshes from Objaverse, Objaverse-XL, and 3D-FUTURE, beginning with around 350,000 textured mesh models. These are split into UV islands, overlapping or overly fragmented cases are removed, and roughly 300,000 independent islands are retained. Very large islands are further filtered for training stability and memory efficiency, with the paper noting that over 97% of islands have fewer than 500 vertices. The final dataset contains about 15,000 objects and 200k UV islands (Chen et al., 25 Sep 2025).

The selection protocol is explicitly style-oriented. Islands with meaningful semantics and organized layouts are manually selected. Blender’s Ministretch is used to compare candidate islands against original UVs with SSIM, and islands with scores between 0.5 and 0.8 are selected as high-quality, manually adjusted cases (Chen et al., 25 Sep 2025). This is significant because the supervision signal is not merely any valid UV map, but a filtered subset intended to encode artist-edited layout priors.

For evaluation, the paper uses ArtUV-200K and the FAM benchmark. On ArtUV-200K, the parameterization stage is compared against Blender, Maya, and 3DsMax using 100 diverse 3D models with artist-marked seams so that seam quality does not confound the comparison. Reported metrics are mesh distortion, UV utilization, and an “Artist-Level” user-study score. On the FAM benchmark, the full pipeline is compared to XAtlas, Nuvo, and FAM using distortion, runtime, and fragment count (Chen et al., 25 Sep 2025).

The unsupervised semantic-visibility ArtUV uses a different evaluation suite. For semantic-aware UV mapping it reports Hamming distance, Rand Index, conformality, equiareality, and inference time, using SAMesh outputs as a reference segmentation. For visibility-aware UV mapping it reports mean seam AO, conformality, equiareality, and inference time, with FlexPara and OptCuts as baselines (Zamani et al., 29 Sep 2025). This divergence in evaluation protocols reflects a deeper methodological split: one ArtUV emphasizes imitation of artist-authored UV style, while the other emphasizes semantic partition quality and seam perceptibility without requiring ground-truth UV atlases.

5. Empirical performance and ablations

On ArtUV-200K, the supervised parameterization stage reports the following results under fixed artist-marked seams: Blender achieves distortion 9.85, utilization 62.74%, and artist-level 3.34; Maya achieves distortion 9.66, utilization 67.53%, and artist-level 1.32; 3DsMax achieves distortion 11.88, utilization 67.01%, and artist-level 1.53; artist-manual UVs achieve distortion 10.90, utilization 70.08%, and artist-level 4.12; and ArtUV reports distortion 9.52, utilization 72.57%, and artist-level 4.22 (Chen et al., 25 Sep 2025). The paper therefore presents ArtUV as simultaneously achieving the highest UV utilization, low distortion, and the top artist-level score in this benchmark.

On the FAM benchmark, the full pipeline reports XAtlas at distortion 9.44, runtime 80.4 s, and 1292 fragments; Nuvo at distortion 32.24, runtime 2925.8 s, and 1 fragment; FAM at distortion 76.28, runtime 5656.3 s, and 1 fragment; and ArtUV at distortion 8.91, runtime 36 s, and 14 fragments (Chen et al., 25 Sep 2025). The paper interprets these results as showing a compromise distinct from both over-fragmenting classical atlases and heavily optimized but impractical single-island solutions. This is consistent with Nuvo’s own design goal of reducing fragmentation on unruly geometry through a continuous neural UV mapping over visible points rather than mesh-vertex parameterization (Srinivasan et al., 2023).

The ablation study is highly diagnostic. Removing silhouette loss harms boundary alignment and utilization. Removing overlap loss causes about 29% overlap and messy maps. Removing distortion loss increases distortion from 9.52 to 10.56 (Chen et al., 25 Sep 2025). These results support the paper’s claim that “artist-style” quality is multi-factorial: boundary regularity, validity, and geometric plausibility are controlled by distinct terms rather than by a single scalar energy.

The unsupervised semantic-visibility ArtUV reports different but related gains. Its semantic-aware pipeline reduces Hamming distance to 0.3188 and improves Rand Index to 0.8151, outperforming xatlas, Blender, Maya, and FlexPara in part alignment; its visibility-aware branch reduces mean seam AO to 0.6065, compared with 0.7855 for OptCuts and 0.8604 for FlexPara, while keeping conformality and equiareality competitive (Zamani et al., 29 Sep 2025). These results indicate that the broader ArtUV agenda is empirically supported from both supervised and unsupervised directions.

6. Relation to adjacent UV research, limitations, and significance

ArtUV sits within a wider movement toward UV-space methods that foreground editability and downstream utility. Nuvo addresses a different failure mode: classical UV algorithms applied to NeRF-derived or otherwise ill-behaved geometry can produce atlases that are too fragmented to be useful, so Nuvo replaces mesh-vertex UV mapping with a continuous neural field defined over visible points and optimized with cycle-consistency, clustering, distortion, and texture-oriented losses (Srinivasan et al., 2023). DreamUV addresses yet another aspect of the same problem by formulating artist-like UV unwrapping as a conditional generative Flow Matching problem over a distribution of professionally authored UV layouts, with boundary-aware weighting and Model-in-the-Loop Finetuning; however, it operates on pre-cut meshes and does not address UV packing, so seam prediction and full atlas construction remain separate problems (Ruan et al., 21 Jun 2026).

The significance of this UV-centric orientation extends beyond parameterization itself. SemUV, for example, performs semantic editing directly in the UV texture space of virtual human heads by training StyleGANv2-ADA on FFHQ-UV and manipulating age, gender, and facial hair via linear semantic boundaries in latent space. Its motivation is precisely that edits performed in UV space remain view-consistent after wrapping the edited texture onto a 3D mesh and integrate cleanly into graphics pipelines such as Blender or PyTorch3D (Mukherjee et al., 2024). This provides a concrete downstream rationale for ArtUV’s insistence on clean, semantically coherent, 2D-editable atlases: the quality of the UV parameterization directly affects the usability of later texture operations.

The supervised ArtUV paper states two explicit limitations. First, the system is highly sensitive to seam quality: if surface cutting is incomplete or inaccurate, the initial UV can be badly distorted and the refinement stage may not fully recover it, even when final edges remain clean. Second, the system does not currently support UV island reuse, because imperfectly reused islands could cause overlap artifacts and complicate training (Chen et al., 25 Sep 2025). The unsupervised ArtUV adds further practical caveats: its Shape Diameter Function partitioning is heuristic, its packing is deliberately simple and grid-based rather than space-optimal, ambient occlusion is only a proxy for human perceptibility, and some variants report about 15 s inference time, slower than sub-second or few-second classical baselines in some cases (Zamani et al., 29 Sep 2025).

Taken together, these papers suggest a technical taxonomy of ArtUV-like systems. One branch learns to imitate artist-authored UV layouts through seam prediction and residual refinement toward curated targets (Chen et al., 25 Sep 2025). A second branch retains unsupervised geometry-preserving UV learning but injects semantic partitioning and visibility-aware seam losses (Zamani et al., 29 Sep 2025). A third, adjacent branch treats artist-like UV structure as a generative prior over pre-cut meshes (Ruan et al., 21 Jun 2026). Across these variants, the common theme is stable: UV parameterization is being redefined as a structured authoring problem in which semantic intelligibility, perceptual seam placement, compactness, and editability are first-class optimization targets rather than by-products of low distortion alone.