PaperlayerView exact revision on arXiv

Source-linked AI summary

Strips as Tokens: Artist Mesh Generation with Native UV Segmentation

Rui Xu, Dafei Qin, Kaichun Qiao, Qiujie Dong, Huaijin Pi, Qixuan Zhang, Longwen Zhang, Lan Xu, Jingyi Yu, Wenping Wang, Taku Komura

arXiv:2604.09132v2cs.CVcs.CGcs.GR

TL;DR

Existing autoregressive mesh tokenizations struggle to preserve the efficient sequences, continuous edge flow, and structural regularity required by professional artist meshes. SATO uses triangle-strip-inspired serialization with native UV boundaries and unified triangle/quad decoding, and reports consistent improvements in geometric fidelity, structural coherence, and UV-aware generation.

  • Problem

    Existing autoregressive tokenizations can produce inefficient sequences or disrupt the continuous edge flow and structural regularity needed for professional artist meshes.

  • Method

    SATO serializes meshes as connected strip-based face chains, encodes UV island boundaries with dedicated tokens, and decodes one sequence as either triangles or quads.

  • Results

    SATO consistently improves geometric fidelity, structural coherence, and UV-aware generation over prior methods across diverse datasets and tasks.

  • Takeaways & Limitations

    The unified representation transfers structural priors from large-scale triangle data to quad meshes while supporting artist-aligned topology and UV chart partitions.

  • Takeaways & Limitations

    Quad quality is bounded by the scale and consistency of available high-quality quad-mesh datasets.

Abstract

from arXiv · show

Recent advancements in autoregressive transformers have demonstrated remarkable potential for generating artist-quality meshes. However, the token ordering strategies employed by existing methods typically fail to meet professional artist standards, where coordinate-based sorting yields inefficiently long sequences, and patch-based heuristics disrupt the continuous edge flow and structural regularity essential for high-quality modeling. To address these limitations, we propose Strips as Tokens (SATO), a novel framework with a token ordering strategy inspired by triangle strips. By constructing the sequence as a connected chain of faces that explicitly encodes UV boundaries, our method naturally preserves the organized edge flow and semantic layout characteristic of artist-created meshes. A key advantage of this formulation is its unified representation, enabling the same token sequence to be decoded into either a triangle or quadrilateral mesh. This flexibility facilitates joint training on both data types: large-scale triangle data provides fundamental structural priors, while high-quality quad data enhances the geometric regularity of the outputs. Extensive experiments demonstrate that SATO consistently outperforms prior methods in terms of geometric quality, structural coherence, and UV segmentation. Project page: https://ruixu.me/html/SATO/index.html

1 Introduction

Professional artist meshes require accurate geometry alongside regular topology and semantic UV layouts, but existing autoregressive tokenizations do not inherently preserve these structures. SATO addresses this with strip-based ordering, unified triangle/quad decoding, native UV segmentation, and consistent improvements over prior methods.

  • Artist meshes support surface editing, connectivity and edge-flow control, deformation, simulation, and texture mapping.
  • Professional mesh generation must combine high-fidelity geometry with clean edge flow and UV layouts compatible with animation and rendering pipelines.
  • SATO orders vertices as connected face chains whose consecutive faces share an edge, aligning token structure with organized artist-mesh edge flow.
  • The same unified vertex sequence decodes into triangle or quadrilateral meshes, allowing triangle data to provide structural priors and quad data to enhance geometric regularity.
  • SATO consistently improves geometric fidelity, structural coherence, and UV-aware generation across diverse datasets and tasks.
  • Dedicated segmentation tokens encode UV island boundaries directly in the sequence, enabling explicit semantic partitioning without sacrificing compression efficiency.

2 Related Work

Prior 3D-generation methods improve geometry, scalability, and UV processing but often leave production topology and semantic layout to separate stages. Existing autoregressive mesh methods remain largely triangle-centric, motivating representations that encode higher-order structure directly.

  • Implicit and hybrid pipelines generate detailed textured geometry but often produce dense meshes requiring post-processing for lightweight artist-style assets.
  • Autoregressive mesh methods serialize meshes into discrete sequences, with successive work targeting fidelity, scale, compression, adjacency, and decoding efficiency.
  • Most existing tokenizations remain triangle-centric, leaving continuous surface runs, stable edge flow, and coherent region growth to many local decisions.
  • Quad-dominant meshes are favored in production for regular edge flow and deformation, but direct generation requires higher-order consistency beyond local triangulation.
  • Production systems commonly treat UV unwrapping and seam placement as downstream or multi-stage processes rather than integrated generation constraints.

3 Preliminaries

Triangle strips compactly represent connected triangles through ordered vertices, while mesh generation converts irregular geometry into sequences for autoregressive prediction. SATO extends this sequence framework to unified mesh decoding conditioned on point clouds.

  • 3.1 Triangle Strips: A triangle strip stores an ordered vertex sequence that implicitly defines m−2 connected triangles.
  • 3.1 Triangle Strips: Consecutive strip triangles share an edge, so each new triangle introduces only one vertex index and improves storage efficiency.
  • 3.1 Triangle Strips: Alternating vertex order flips neighboring face orientation, requiring reordering or a parity toggle during decoding.
  • 3.1 Triangle Strips: Artists’ incremental boundary extension produces coherent strips with stable local connectivity and clear sequential order.
  • 3.2 Autoregressive Mesh Generation Framework: Mesh tokenization serializes geometric and topological data into a discrete sequence that bridges irregular 3D structures and standard sequence models.
  • 3.2 Autoregressive Mesh Generation Framework: Conditioned on a point cloud, an autoregressive Transformer predicts each next token from preceding context.

4 Method

SATO uses hierarchical quantization, strip serialization with UV transition markers, adaptive triangle/quad decoding, and staged training from triangles to high-quality quads. Together, these components form a unified artist-style mesh-generation framework.

  • SATO maps 3D coordinates into a compact discrete vocabulary through hierarchical geometry quantization.
  • Strip serialization converts meshes into contiguous vertex streams with embedded markers for UV transitions.
  • A multi-topology interpretation protocol adaptively decodes the recovered sequence as either triangle or quadrilateral meshes.
  • SATO trains through large-scale triangle pretraining followed by fine-tuning on high-quality quad meshes.

4.1 Hierarchical Geometry Quantization

SATO represents artist meshes as vertices and faces, then discretizes vertex coordinates through a three-level hierarchical quantization scheme. The resulting coordinate tuples encode geometry but remain unordered and detached from topology.

  • An artist mesh M is represented by vertices V and faces F, with each face an ordered sequence of vertex indices.
  • Triangle and quadrilateral faces are distinguished by degree |f| equal to 3 or 4, respectively.
  • Vertex coordinates are quantized on a 512^3 voxel grid after normalizing the mesh into a unit cube.
  • Each vertex becomes a hierarchical tuple (c_1,c_2,c_3) at 4^3, 8^3, and 16^3 resolution levels.c_1 identifies the coarsest grid cell, while c_2 and c_3 specify progressively finer local positions.
  • After quantization, the geometry is fully discretized, but the coordinate tuples remain unordered and detached from the mesh faces.

4.2 Strip-based Serialization

SATO serializes meshes as connected strips of adjacent faces, embedding strip and UV-island transitions directly into the token stream. This ordering preserves edge-flow continuity, supports unified triangle/quad decoding, and reduces transition overhead compared with patch-based orderings.

  • Strip extraction: A strip is a connected face sequence in which consecutive faces share an edge, producing a coherent vertex stream aligned with artist-mesh edge flow.The extraction procedure grows strips across adjacent unvisited faces from a seed until reaching a boundary or visited region.
  • Unified triangle/quad representation: SATO uses stride δ=1 for triangle strips and δ=2 for quad strips, allowing both mesh types to follow the same grow-by-appending traversal.Each triangle step adds one vertex, while each quad step adds two vertices to induce the next face.
  • Strip transitions: Specialized coarse-level tokens mark the first vertex of each strip, embedding strip boundaries without adding separate delimiter tokens or sequence length.These structural tokens reset prefix sharing so topological transitions remain explicit and unambiguous.
  • UV segmentation: UV segmentation is encoded by partitioning faces into UV islands, traversing each island completely before transitioning, and inserting tokens that mark island completion and the next chart.The model learns chart partitioning rather than UV coordinates, which are assigned later by a standard unwrapping algorithm.
  • Sequence compression: Prefix sharing omits repeated coarse coordinate prefixes between consecutive vertices, while absolute strip and UV tokens force synchronization resets.On the test set, token frequencies are c1: 20.7%, c2: 35.0%, and c3: 44.3%, indicating compression of many vertices to one or two tokens.

4.3 Topology-Specific Decoding

SATO deterministically decodes one token stream into triangle or quadrilateral meshes using structural markers and an adjustable vertex stride. Strip and UV-island markers reset the decoding state at segment boundaries.

  • Decoding protocol: The decoder reconstructs hierarchical vertex coordinates and uses structural tokens to manage strip and UV-island boundaries.C₁ terminates a strip, while Cᵤᵥ₁ marks a transition between disjoint UV islands; either marker resets the coordinate cache and topological frontier.
  • Topology-specific recovery: An adjustable stride δ∈{1, 2} enables multi-topology recovery from the same vertex ordering.The decoder applies different face-construction rules depending on whether it operates in triangle or quadrilateral mode.
  • Triangle mode: With δ=1, each successive vertex completes a triangle with the two preceding vertices.The triangle face is formed as f_i=(v_i,v_i+1,v_i+2).
  • Quadrilateral mode: With δ=2, the decoder processes vertices in pairs to recover quadrilateral topology.The supplied passage identifies pairwise processing as the quadrilateral decoding rule.
  • Connectivity: Vertices from different strips sharing quantized coordinates within a UV region are welded during decoding to ensure connectivity.The passage also notes a failure mode involving consecutive structural tokens, but reports that trained models have not exhibited it.

4.4 Training with SATO

SATO trains progressively from triangle geometry to UV segmentation and then quadrilateral generation. This staged design uses abundant triangle data for priors and transfers them to scarcer UV and quad tasks.

  • Training strategy: SATO’s training pipeline has three stages: triangle-mesh pretraining, UV-segmentation post-training, and quad-mesh fine-tuning.The model is trained sequentially rather than learning all objectives from scratch.
  • Data preparation: The curated training corpus combines multiple mesh datasets after filtering invalid, fragmented, and unsuitable models.Filtering includes removing non-manifold models, merging duplicate vertices, and restricting face counts and vertex-to-face ratios.
  • Unified tasks: The framework supports triangular mesh generation, UV segmentation generation, and quadrilateral mesh generation within one model.The output gallery presents these three tasks in a single framework.
  • Stage I: Triangle Mesh Pretraining: Stage I trains the backbone and base tokenizer on large triangle datasets to establish geometric and conditioning priors.These include local strip continuation patterns and alignment between mesh tokens and conditioning point clouds.
  • Stage II: UV Segmentation Post-Training: UV post-training initializes from the pretrained triangle model to accelerate convergence and improve segmentation performance.This retains learned geometric and conditioning alignment while introducing UV tokens and inter-island transition rules.
  • Stage III: Quad Mesh Fine-tuning: Quad fine-tuning transfers triangle-domain priors to quadrilateral generation, which requires relatively little quad data.The representation’s compatibility makes training a quad generator from scratch unnecessary at scale.

5 Experimental Results

Across triangle generation and UV segmentation experiments, SATO is evaluated against public baselines on held-out shapes and receives favorable qualitative and user-study results. Its UV outputs are also reported to support practical texturing and lower distortion.

  • Experimental scope: SATO supports triangle generation, UV segmentation, and quadrilateral generation within a single framework.Representative outputs are shown for all three tasks.
  • Evaluation setup: The evaluation uses four geometric metrics—NC, CD, HD, and F1—on a consistent 250-shape test set from three datasets.The test shapes are excluded from training, and all methods receive the same lightweight post-processing for fair evaluation.
  • Triangle Mesh Generation: SATO consistently outperforms four public baselines across multiple geometric metrics on ShapeNet, Thingi10K, and Objaverse.The reported baselines are MeshAnythingV2, BPT, TreeMeshGPT, and DeepMesh.
  • Triangle Mesh Generation: Artists consistently prefer SATO’s outputs in the triangle-mesh user study using ranking-based scores.Participants judged regularity, artist-likeness, geometric fidelity, and shape consistency holistically.
  • UV Segmentation: SATO produces cleaner, more organized UV segmentation than PartUV-related comparisons and yields lower distortion across four UV metrics.The reported segmentation boundaries better align with geometric features and produce more regular islands.
  • UV Segmentation: Compared with MeshMosaic, SATO mitigates overly long seams and produces cleaner, more regular segmentation.The comparison attributes MeshMosaic’s visible seams and asymmetry artifacts to reliance on precomputed part boundaries.
  • UV Segmentation: The generated UV layouts support targeted texture painting by separating components into well-defined islands.The paper describes this as practical downstream utility for artists.

5.3 Quad Mesh Generation

SATO generates quadrilateral meshes by changing the detokenizer rather than the model architecture. Its quad outputs retain geometric fidelity while offering compact, structured layouts that users prefer to remeshing and reconstruction baselines.

  • Quad generation: SATO generates quadrilateral meshes by switching the detokenizer without changing the model architecture.The quad detokenizer merges adjacent triangle pairs from the same token sequence.
  • Comparison with triangle generators: Compared with prior triangle-mesh generation models, SATO more consistently produces high-quality, stable quadrilateral meshes and native UV segmentation.The qualitative comparison covers diverse shapes.
  • Geometric fidelity: Quad outputs have nearly identical geometric fidelity to corresponding triangle outputs.This follows from decoding both topologies from the same geometric token sequence.
  • Comparison with remeshing methods: Against remeshing and reconstruction methods, SATO produces compact quad layouts with high quad utilization and alignment to salient feature lines.The baselines often struggle to achieve high quad utilization, low face count, and feature-line alignment simultaneously.
  • Geometric evaluation: Despite generating from a point cloud rather than ground-truth geometry, SATO achieves competitive or superior geometric scores in the reported comparison.Remeshing methods have near-identical fidelity because they operate directly on ground-truth geometry.
  • User study: Users clearly prefer SATO’s quadrilateral outputs over the five reconstruction and remeshing baselines.The user study reports ranking-based scores in the range [0, 3].

5.4 Ablation Studies

Ablation studies show that SATO’s tokenizer learns UV segmentation and reconstructs meshes efficiently, while staged UV training and quad fine-tuning improve artist-aligned outputs. The framework also supports image- and text-conditioned generation, diverse outputs, and native UV segmentation.

  • Tokenizer ablation: SATO converges faster than DeepMesh’s tokenizer and reaches near-perfect reconstruction with clean, well-structured UV segmentation earlier in training.Both settings use the same model architecture and training hyperparameters; tokenizer choice is the only difference.
  • Tokenizer ablation: 24K versus 20K tokens means SATO uses about 85% of DeepMesh’s token length for the teapot encoding.The reduction is attributed largely to fewer patch/strip transitions: 0.9K versus 1.6K.
  • UV training strategies: UV supervision is applied after triangle pretraining because imposing it from the beginning can hinder fine-grained geometric learning and input alignment.The staged strategy first uses triangle data without UV segmentation, then post-trains on UV-segmented data.
  • Quad-mesh fine-tuning: Quad-mesh fine-tuning produces cleaner, more quad-like routing and improves the quality and artist alignment of generated meshes.It encourages neater mesh routing and increases the prevalence of well-shaped, often right-angled triangles.
  • Image and text generation: SATO remeshes generated shapes from image or text prompts into lightweight triangular or quadrilateral meshes with native UV segmentation.The method is used after CLAY-based 3D generation and produces outputs directly usable in practice.
  • Diversity: Conditioned on the same input, SATO generates diverse mesh geometries and UV segmentations whose charts remain clean, structured, and often symmetric.The diversity results are reported for both geometry and UV layout.

6 Limitations and Future Work

SATO’s unified tri/quad design has three main limitations: occasional triangle degeneration in quad decoding, quad quality bounded by available training data, and irregular routing on near-spherical shapes.

  • Quad decoding: Quad decoding is predominantly quad-dominant, but odd-length strips or repeated vertices can produce local degenerate triangles.The authors describe these cases as structurally well-defined and propose improved data or lightweight post-processing as future mitigations.
  • Quad-data limitation: Attainable quad quality is bounded by the scale and consistency of available high-quality quad-mesh datasets.The authors contrast this unified approach with quad-only methods optimized specifically for quad generation.
  • Shape-dependent routing: Near-spherical shapes can exhibit less regular edge routing because high-quality spherical exemplars are scarce in existing quad corpora.The authors expect this gap to narrow with richer quad datasets and stronger shape-adaptive routing priors.

7 Conclusion

SATO combines strip-based tokenization, native UV boundary encoding, and unified triangle/quad interpretation to generate diverse, high-fidelity artist meshes with strong topological quality.

  • Conclusion: SATO follows artist-mesh edge flow, directly encodes UV island boundaries, and supports mixed-data training through unified triangle/quad decoding.The same sequence format transfers and strengthens priors across mesh formats.
  • Conclusion: Extensive experiments show diverse, high-fidelity meshes with stronger topological quality than competitive baselines.The authors highlight practical potential for downstream content-creation pipelines.
Loading 2604.09132v2…