SeamGPT: Auto-Regressive Surface Cutting

• SeamGPT is an auto-regressive model that reframes 3D mesh seam prediction as a sequence modeling task for artist-like UV unwrapping.
• It uses a transformer-based decoder with a point cloud encoder to generate quantized coordinate tokens, reducing fragmentation and distortion.
• SeamGPT integrates into the ArtUV pipeline to deliver semantically organized charts, demonstrating superior UV map quality on diverse meshes.

Searching arXiv for papers on SeamGPT and closely related UV unwrapping work. SeamGPT is an auto-regressive seam prediction model for 3D meshes that formulates surface cutting as a sequence modeling problem. In computer graphics, seams are the cut curves that partition a mesh into charts suitable for UV parameterization, texture mapping, and related decomposition tasks. SeamGPT was introduced as an “Auto-Regressive Surface Cutting” model and is used in ArtUV as the learned surface-cutting stage of a two-stage, top-down UV unwrapping pipeline, where it provides the semantically meaningful cutting seams on which artist-style UV parameterization depends (Li et al., 22 Jun 2025, Chen et al., 25 Sep 2025).

1. Definition and problem formulation

SeamGPT addresses the classical surface-cutting problem on triangular meshes by predicting seam curves that partition a surface into usable charts. Within ArtUV, given a triangular mesh

$$M = (V, F), \quad V \in \mathbb{R}^{N \times 3}, \quad F \in \mathbb{R}^{M \times 3},$$

the task is to identify a subset of vertices

$V_c \subseteq V$

such that these vertices, connected via mesh edges, form seam curves partitioning the surface into charts that can later be flattened to UV space. In the broader SeamGPT framing, surface cutting is the process of turning a general 3D mesh into a set of topological disks or charts by inserting cut curves on the surface (Li et al., 22 Jun 2025, Chen et al., 25 Sep 2025).

The problem is not merely geometric. The stated deficiencies of existing automatic UV unwrapping and seam-generation methods include time consumption, fragmentation, lack of semanticity, irregular UV islands, and the tendency to optimize purely geometric criteria without encoding global shape context. SeamGPT is intended to emulate a professional artist’s first step in UV unwrapping: deciding where to cut the surface so that it breaks into logical, usable charts. Its stated targets include semantic seam placement, reduced fragmentation, and applicability to both artist-created meshes and AI-generated meshes (Chen et al., 25 Sep 2025).

A central conceptual move is the conversion of seam placement into sequence prediction. Seam curves are encoded as coordinate tokens, and the model autoregressively predicts these tokens. In this formulation,

$$p(\tau \mid c) = \prod_{t=1}^T p(\tau_t \mid \tau_{<t}, c),$$

where $c$ is a shape-conditioned representation and $\tau$ is the seam-token sequence. This recasts seam generation as next-token prediction rather than local geometric optimization (Li et al., 22 Jun 2025).

2. Architecture and tokenized seam representation

SeamGPT is described as a GPT-like autoregressive transformer specialized for meshes rather than a text-based GPT in the LLM sense. Its architecture combines a point cloud encoder, used to embed mesh geometry into a latent shape condition, with an hourglass-style autoregressive Transformer decoder inspired by Meshtron. The decoder uses causal masking so that each token is generated conditioned only on previously generated seam tokens and the shape condition (Chen et al., 25 Sep 2025).

Its input representation is built from structured point sampling on the mesh. The reported configuration samples 30,720 points on vertices and 30,720 points along edges, for a total of 61,440 points. This sampling is designed to capture both vertex locations and edge structures, and in the more detailed surface-cutting description the edge samples are stated to be proportional to edge length. The point cloud is then encoded into a latent shape embedding or “shape condition” that conditions decoding (Li et al., 22 Jun 2025, Chen et al., 25 Sep 2025).

Seam geometry is represented through quantized coordinate tokens. ArtUV reports that seam vertices are spatially sorted and coordinates are quantized into discrete tokens using 10-bit encoding per coordinate, with a seam segment represented by six consecutive tokens. The decoder outputs coordinate tokens, which are decoded back into continuous coordinates and then mapped onto the mesh by nearest-point projection to the nearest mesh vertex or edge point; the resulting points are connected along edges to form seam curves on the mesh surface (Chen et al., 25 Sep 2025).

The decoder is a large long-sequence transformer. The appendix-reported configuration uses a three-level abstraction structure with depth configuration

$(2,\ (4, 12, 4),\ 2),$

hidden dimension 1,536 per Transformer block, 16 attention heads, and positional encoding for sequence lengths up to 36,864 tokens. A separate description of the SeamGPT training pipeline states that a length embedding is concatenated to the shape embedding, with a maximum truncated training length of 27,000 tokens to stabilize long-sequence generation. This suggests explicit control over seam-sequence length in addition to geometric conditioning (Li et al., 22 Jun 2025, Chen et al., 25 Sep 2025).

3. Training regime, supervision, and data

SeamGPT is trained on data that includes artist-created meshes and AI-generated meshes. The intended learning signal is not explicit semantic annotation such as “arm” or “torso,” but the seam patterns embodied in meshes whose cutting decisions reflect human practice. In the surface-cutting paper, the main objective is the token-level auto-regressive cross-entropy loss

$$\mathcal{L}_{\text{CE}} = - \sum_{t=1}^T \log p_\theta(\tau_t \mid \tau_{<t}, c),$$

with the model optimized to predict the next seam token conditioned on prior tokens and the shape representation (Li et al., 22 Jun 2025).

Two training descriptions are reported. ArtUV states that the complete SeamGPT model was reproduced after consultation with the SeamGPT authors, and reports a setup of 64 H20 GPUs (96 GB each), 200k training steps, fixed learning rate $1 \times 10^{-4}$, gradient clipping at 0.5, batch size 128, and data augmentation consisting of random scaling in $[0.95, 1.05]$, random vertex jitter with noise level 0.01, and random rotations. The surface-cutting description separately reports Adam optimization with fixed learning rate $10^{-4}$, no weight decay, gradient clipping at 0.5, batch size 2, and a one-step warm-up phase, together with truncated-mode training for long sequences (Chen et al., 25 Sep 2025, Li et al., 22 Jun 2025).

The reported augmentations are intended to improve robustness to scale, noise, and orientation. The data regime itself is described as diverse, spanning manifold and non-manifold meshes, artist-created models, and 3D-scanned models. In the broader SeamGPT framing, the model is presented as learning robust priors for seam placement across varied topology and geometry rather than solving each mesh independently with hand-crafted heuristics (Li et al., 22 Jun 2025).

4. SeamGPT inside ArtUV

ArtUV is explicitly a two-stage, learning-based top-down UV unwrapping pipeline. Stage 1 is surface seam prediction by SeamGPT. Stage 2 is artist-style UV parameterization by an Auto-Encoder that refines an optimization-based UV into an artist-style map. In the first stage, SeamGPT predicts seams on the input mesh, and the mesh is segmented along these seams into charts. In the second stage, each chart is unwrapped quickly with a variant of Ministretch-Unwrap combining techniques such as SLIM, SCAF, and LSCM to produce a rough, low-distortion UV; the rough UV together with mesh features is then refined by the ArtUV Auto-Encoder into a UV map with clean boundaries, good space utilization, and preserved topology (Chen et al., 25 Sep 2025).

The parameterization stage is formulated in terms of an initial UV $V_c \subseteq V$0, mesh features including vertices $V_c \subseteq V$1, faces $V_c \subseteq V$2, normals $V_c \subseteq V$3, vertex degrees $V_c \subseteq V$4, and curvature $V_c \subseteq V$5, and a predicted per-vertex UV offset $V_c \subseteq V$6. The final prediction is reported as

$V_c \subseteq V$7

Within this design, SeamGPT determines the cutting geometry and semantics that condition all downstream UV processing (Chen et al., 25 Sep 2025).

The notion of “semantically meaningful cutting seams” in ArtUV includes seams that follow natural object part boundaries, avoid splitting important regions excessively, place texture discontinuities in less visually prominent areas, and produce charts that correspond to intuitive editing regions. ArtUV states that SeamGPT does not explicitly use semantic labels, and that semantics are instead learned implicitly from artist-created meshes and AI-generated meshes through the association between shape configurations and appropriate seam placements (Chen et al., 25 Sep 2025).

ArtUV also makes explicit that seam quality is structurally decisive. Poor seam placement affects chart boundaries and chart topology, and can cause severe distortions during UV initialization. The paper states that “The method’s performance is highly sensitive to the quality of surface cutting. Incomplete or inaccurate seams may cause severe distortions during UV initialization…”. In that sense, SeamGPT provides the structural backbone of the pipeline, while the Auto-Encoder performs refinement rather than topological repair (Chen et al., 25 Sep 2025).

5. Evaluation and comparative performance

In ArtUV’s FAM benchmark evaluation of the full pipeline, including SeamGPT, the reported face conformal energy distortion values are 9.44 for XAtlas, 32.24 for Nuvo, 76.28 for FAM, and 8.91 for ArtUV with SeamGPT. Reported runtimes are 80.4 s for XAtlas, 2925.8 s for Nuvo, 5656.3 s for FAM, and 36 s for ArtUV. Reported fragment counts are 1292 for XAtlas, 1 for Nuvo, 1 for FAM, and 14 for ArtUV. The qualitative interpretation given in the paper is that XAtlas produces extreme fragmentation, Nuvo and FAM yield disordered topology or overlapping unusable UV maps, and ArtUV produces compact, semantically organized charts with clear islands (Chen et al., 25 Sep 2025).

The low fragment count is particularly significant for understanding SeamGPT’s role. ArtUV explicitly connects the full-pipeline result of 14 fragments to seams that yield large, semantically meaningful charts rather than hundreds or thousands of tiny fragments. This is the most concrete numerical evidence, in the provided material, that SeamGPT’s seam-selection stage changes the topology of the resulting atlas in a way aligned with artist workflows (Chen et al., 25 Sep 2025).

A second benchmark, ArtUV-200K, isolates the parameterization stage by using artist-marked seams shared across methods. In that setting SeamGPT is not used, but the results clarify the upper bound ArtUV can approach when seam quality is high. Reported distortion values are 9.85 for Blender, 9.66 for Maya, 11.88 for 3DsMax, 10.90 for Artist-manual, and 9.52 for ArtUV; reported UV utilization values are 62.74, 67.53, 67.01, 70.08, and 72.57, respectively; Artist-Level scores are 4.12 for Artist-manual and 4.22 for ArtUV. The paper interprets these results as evidence that good seams enable the parameterization stage to produce better-than-artist UV in distortion and space utilization, and that SeamGPT is intended to automate comparable seam quality on arbitrary meshes (Chen et al., 25 Sep 2025).

The surface-cutting paper adds a broader claim that SeamGPT achieves exceptional performance on UV unwrapping benchmarks containing both manifold and non-manifold meshes, including artist-created and 3D-scanned models, and that it enhances existing 3D segmentation tools by providing clean boundaries for part decomposition. The detailed excerpt does not provide a numerical table for these comparisons, but it does state lower mean conformal distortion, fewer chart fragments, and higher segmentation accuracy when SeamGPT seams are used as structural cues (Li et al., 22 Jun 2025).

6. Relation to prior methods, limits, and broader significance

SeamGPT differs from classical seam strategies in both representation and optimization philosophy. Traditional seam-generation methods include optimization-based joint cutting and parameterization, such as OptCuts, and heuristic seam placement based on curvature or angle thresholds. These methods are described as geometric rather than semantic, and as prone to fragmented or non-intuitive charts. SeamGPT instead learns from human seam decisions and AI-generated models, uses a global autoregressive transformer rather than local geometric rules, and focuses specifically on seam prediction rather than end-to-end UV generation (Li et al., 22 Jun 2025, Chen et al., 25 Sep 2025).

The “GPT” designation refers to autoregressive, token-based generation and transformer architecture, not to natural-language interaction. ArtUV explicitly notes that SeamGPT is not a text-based GPT in the LLM sense: it uses multi-head attention, causal masks, and next-token prediction over seam tokens, but has no natural-language input or output. This nomenclature emphasizes the sequence-modeling formulation of surface cutting rather than linguistic functionality (Chen et al., 25 Sep 2025).

Its limitations are correspondingly structural. ArtUV directly identifies sensitivity to seam quality as the central limitation: incomplete or inaccurate seams can cause severe internal distortions during UV initialization, and the Auto-Encoder cannot fully repair structural issues introduced at the cutting stage. ArtUV also notes that UV island reuse is not yet supported, and that seam placement interacts with reuse because misaligned seams between similar meshes hinder reuse and can cause overlap. Proposed future directions include improving seam quality and stability, secondary segmentation of high-distortion areas, more advanced seam refinement or multi-stage SeamGPT usage, and integrating UV island reuse through similarity-based merging of optimized islands across models (Chen et al., 25 Sep 2025).

A plausible implication is that SeamGPT exemplifies a more general seam-centric view of geometric structure: seams can function not only as UV-cut boundaries but also as compact structural priors. A separate garment-manipulation paper, though not about SeamGPT itself, maps seam observations to a topology-encoded structural skeleton graph for garment state estimation, indicating that seam-based representations can support control-oriented reasoning beyond UV unwrapping (Huang et al., 13 Jun 2026). Within graphics, however, SeamGPT’s specific contribution is the recasting of surface cutting as long-horizon conditional sequence generation, with seams treated as learned, semantically informed geometric programs rather than purely local optimization artifacts (Li et al., 22 Jun 2025).