Organizations: The University of Hong Kong, Deemos Technology Co., Ltd. , China · ShanghaiTech University, Deemos Technology Co., Ltd., China · Shandong University, China · The University of Hong Kong, China · ShanghaiTech University, China · Texas A&M University, USA
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
Figures & tables
Figure 1. Strips as Tokens (SATO) enables unified, high-quality artist mesh generation with native UV segmentation. Our strip-based tokenizer supports both triangle (left) and quad (right) meshes without retraining and automatically segments UV charts (side) during autoregressive generation.
Figure 2. Artist meshes differ markedly from geometry-processed ones. Here we show quadrilateral and triangular meshes constructed by artists, as well as meshes created using geometric processing methods (such as Delaunay-style remeshing ( Xu et al., 2024 ) and Marching Cubes ( Lorensen and Cline, 1998 ) ).
Figure 3
Figure 3. The Pipeline of SATO . SATO uses a strip-based tokenizer to encode/decode both triangle and quad meshes as a unified discrete sequence. Conditioned on an input point cloud, a learnable point-cloud encoder cross-attends to the core Hourglass Transformer, which autoregressively generates token sequences that are decoded into triangle or quad meshes with native UV segmentation.
Figure 5Figure 6
Figure 4. Mesh tokenization with prefix sharing. We use a three-level hierarchical coordinate (c1,c2,c3) with prefix sharing to compress token sequences following DeepMesh ( Zhao et al., 2025 ) , repeated prefixes between consecutive vertices are omitted. To mark a strip transition, we introduce a special top-level token vocabulary c1∗ (red), which is distinct from c1 but serves the same role. Whenever c1∗ appears, the prefix sharing state is reset, starting a new prefix context.
Figure 5. Artist-created meshes with UV chart partitions. We split artist meshes into UV parts and let SATO traverse all triangles within one part before a UV segmentation transition to the next part, enabling native UV segmentation during generation.
Figure 6. Unified representation of triangle and quad using strips. Triangle strips may locally “turn” under edge flips (a). In contrast, quad strips avoid this ambiguity (b), as each step admits only a single forward direction. Moreover, sequences tokenized from a quad mesh can be decoded into triangles while still preserving high quality (c). Note that the quad token sequence of (b) is totally the same as the triangle token sequence of (c).
Metrics
BPT
DeepMesh
SATO
Vocab Size ↓
40960
4736
4800
Comp Rate ↓
0.228
0.330
0.283
Table 1. Comparison of vocabulary size and average compression rate. The compression rate is computed as the token sequence length divided by (face count × 9).
Figure 7. Different face ordering defined by other methods. BPT ( Weng et al., 2025 ) and DeepMesh ( Zhao et al., 2025 ) traverse local fan-/disk-shaped neighborhoods, i.e., triangles rotate around a vertex, which triggers patch transitions more frequently. In contrast, our strip-based ordering can, in principle, extend arbitrarily long.
Figure 8. The gallery of SATO illustrates our model’s outputs across three tasks. From bottom to top, it shows triangular mesh generation, shape generation with UV segmentation, and quadrilateral mesh generation. SATO supports all three tasks within a single framework and achieves compelling results on each of them.
ShapeNet
Thingi10K
Objaverse
Method
NC↑
CD↓
HD↓
F1↑
NC↑
CD↓
HD↓
F1↑
NC↑
CD↓
HD↓
F1↑
MeshAnythingV2 ( Chen et al., 2024c )
0.911
0.009
0.078
0.361
0.841
0.022
0.168
0.162
0.858
0.016
0.117
0.208
TreeMeshGPT ( Lionar et al., 2025 )
0.840
0.034
0.161
0.439
0.791
0.058
0.228
0.236
0.783
0.057
0.238
0.188
BPT ( Weng et al., 2025 )
0.962
0.003
0.017
0.605
0.874
0.028
0.141
0.248
0.841
0.030
0.137
0.265
DeepMesh ( Zhao et al., 2025 )
0.967
0.004
0.037
0.532
0.853
0.026
0.167
0.157
0.859
0.020
0.120
0.240
SATO
0.975
0.002
0.032
0.807
0.916
0.009
0.154
0.460
0.909
0.009
0.117
0.503
Table 2. Quantitative comparison on ShapeNet ( Chang et al., 2015 ) , Thingi10K ( Zhou and Jacobson, 2016 ) , and Objaverse ( Deitke et al., 2023 ) datasets. The best scores are emphasized in bold with underlining, while the second best scores are highlighted only in bold.
Figure 9. Overview of our test dataset. 250 models selected from the ShapeNet ( Chang et al., 2015 ) , Thingi10K ( Zhou and Jacobson, 2016 ) and Objaverse ( Deitke et al., 2023 ) dataset.
Figure 10. Qualitative comparison with baseline methods across different shapes. Our approach consistently produces high-quality artist meshes with stable structure and clean surface.
MeshAthV2
TreeMeshGPT
BPT
DeepMesh
Ours
Scores
0.18
0.57
1.4
1.17
2.61
Variance
0.27
0.67
0.95
0.93
0.49
Table 3. User study with SOTA methods on triangle mesh generation. Each score is the mean ranking-based score over all participants (range [0,3] ; 1st=3, 2nd=2, 3rd=1, others=0).
Figure 11. Qualitative comparison with PartUV ( Wang et al., 2025c ) . Our method generates an artist mesh (a) together with explicit UV segmentation (b). By applying angle-based UV unwrapping from Blender ( Blender, 2025 ) , we further obtain a high-quality 2D UV layout (c). In contrast, PartUV relies on a PartField ( Liu et al., 2025c ) pre-segmentation pipeline; regardless of whether it is applied to our generated mesh (d) or the ground-truth (GT) mesh (e), its resulting UV charts are consistently less clean and less well-structured than ours.
Figure 12. Gallery of UV unwrapping results using our generated UV segmentation. The shapes are taken from Fig. 8 , and both UV unwrapping and visualization are obtained through the unwrapping algorithm in Blender ( Blender, 2025 ) .
Figure 13. Comparison with MeshMosaic ( Xu et al., 2025 ) . Our method yields cleaner, more regular segmentation and mitigates the issue of overly long seams.
PartUV w/ Our Mesh
PartUV w/ GT Mesh
Ours
Scores
2.04
1.36
2.6
Variance
0.49
0.36
0.38
Table 4. User study with PartUV ( Wang et al., 2025c ) . The scores are calculated based on the rankings and range from [0,3] .
Method
L2 Stretch ( ↓ to 1)
Area Dist. ( ↓ )
Angle Dist. ( ↓ to 1)
Sym. Dirichlet ( ↓ to 4)
PartField
0.921
0.849
1.256
8.283
SATO
0.979
0.562
1.128
5.156
Ground Truth
1.010
0.146
1.138
4.102
Table 5. UV distortion comparison. Our segmentation produces consistently lower UV distortion than PartField ( Liu et al., 2025c ) across all four standard metrics with Blender’s ( Blender, 2025 ) angle-based unwrapping.
Figure 14. Comparison with MeshSilksong ( Song et al., 2025 ) . Our method produces higher-quality meshes and cleaner, more coherent segmentation.
Figure 15. Texture painting with our UV unwrapping. The high-quality UV unwrapping produced by our method makes it easy for artists to paint texture maps.
Figure 16. Qualitative comparison with BPT ( Weng et al., 2025 ) and DeepMesh ( Zhao et al., 2025 ) on diverse shapes. Compared with prior triangle-mesh generation models, our method more consistently generates high-quality quadrilateral meshes, is more stable, and additionally predicts native UV segmentation.
Figure 17. Qualitative comparison with quad remeshing and reconstruction methods. Due to reliance on quadrilateral parameterization, these methods typically struggle to produce highly simplified quad meshes. In contrast, our method can generate meshes with diverse densities and additionally supports native UV segmentation.
IM
QuadriFlow
Quadwild
NeurCross
CrossGen
Ours
Scores
0.12
0.28
1.08
1.48
1.24
1.8
Variance
0.20
0.44
1.23
1.32
1.28
1.28
Table 6. User study with remeshing and reconstruction methods on quad mesh generation. The scores are calculated based on the rankings and range from [0,3] .
Method
NC↑
CD↓
HD↓
F1↑
IM
0.917
0.007
0.052
0.304
QuadriFlow
0.924
0.005
0.080
0.451
QuadWild
0.970
0.001
0.020
0.848
NeurCross
0.968
0.002
0.020
0.846
CrossGen
0.969
0.001
0.020
0.849
Ours
0.971
0.001
0.020
0.857
Table 7. Geometric metrics on the quad mesh from Fig. 17 . Remeshing methods achieve near-identical fidelity to the input since they operate directly on the ground-truth geometry, whereas our method generates the mesh from a point cloud.
Methods
Token Length ↓
Transitions ↓
Time (s) ↓
Training Speed (Steps/s) ↑
Infer (Tokens/s) ↑
DeepMesh
24674
1654
2.061
0.442
∼ 55
Ours
20830
981
0.319
0.488
∼ 58
Table 8. Quantitative comparison with the DeepMesh ( Zhao et al., 2025 ) tokenizer. Under the overfitting setup in Fig. 18 , we report compression and training metrics for both tokenizers. Our method achieves better training speed and a higher compression ratio.
Figure 18. Ablation with the DeepMesh ( Zhao et al., 2025 ) tokenizer. We constructed an overfitting ablation to compare our tokenizer with DeepMesh. Our tokenizer converges faster and is easier for the network to learn, even when augmented with UV segmentation.
Tokenizer
NC↑
CD↓
HD↓
F1↑
DeepMesh
0.908
0.022
0.108
0.455
Ours
0.925
0.014
0.103
0.560
Table 9. Large-scale tokenizer ablation. Both models use the same architecture, data, and training budget ( 64× A800, 200K steps); only the tokenizer differs.
Figure 19. Ablation on UV training strategies with a manually generated unseen test shape. All methods use the same test shape, which is manually created with Rodin ( Zhang et al., 2024a ) from an input text prompt cute bunny astronaut toy , converted into an SDF, and then meshed with marching cubes. This construction guarantees that the shape lies strictly outside the training set. We compare SOTA methods and three training strategies: (a) using a pretrained point cloud VAE with UV segmentation data, (b) training from scratch on UV data, and (c) first training from scratch without UV data followed by post training on UV data. The first two strategies that directly use UV data cannot reliably align with the input shape and usually collapse to an almost random shape.
Figure 20. Ablation on quad-mesh fine-tuning. After quad-mesh fine-tuning, the meshes in the black boxed region become markedly higher quality and more artist-aligned, with cleaner structure and easier downstream editing.
Figure 21. Generation from image and text prompts. By leveraging CLAY ( Zhang et al., 2024a ) for 3D generation, SATO can produce high-quality artist meshes with native UV segmentation from either an input image or a text prompt.
Figure 22. Diversity results. Conditioned on the same input, our model generates diverse meshes and segmentation outcomes, demonstrating strong generative diversity.
Figure 23. Failure cases. Left: Degenerate triangle faces arising from quadrilateral strip decoding under irregular strip configurations. Right: Suboptimal quad layouts limited by the scale and consistency of available high-quality quad-mesh training data.
Autoregressive models have proven effective for triangular mesh generation by transforming it into predicting the next face, converting the non-differentiable connectivity prediction into a sequence of token predictions. However, generating quad-only meshes remains challenging because quad faces impose stricter topological constraints than triangles. The model must not only serialize quad faces into a valid token sequence, but also support unambiguous connectivity reconstruction during de-serialization. To address this, we propose Q-MAGIC, a novel autoregressive model for quad mesh generation. We first introduce a tailored serialization scheme for quad meshes, which requires at least two new vertices for the next face and is less flexible than triangles due to strict wiring constraints. To meet these requirements and reduce serialization length, we propose a singularity-aware serialization that prioritizes wiring from singular points, leading to more coherent traversals. Additionally, we propose a circular relative offset tokenization that encodes geometry mainly via local relative offsets. Built upon this representation, an autoregressive Transformer learns the underlying distribution of quad meshes via next-token prediction, enabling effective quad mesh generation. Extensive experiments demonstrate that Q-MAGIC produces more compact sequences while consistently improving geometric fidelity and mesh quality compared to state-of-the-art methods.
Yuming Zhao, Zezeng Li, Weimin Wang +2
Dalian University of Technology, China · École Centrale de Lyon, France
Prior artisan mesh generation works largely predict face tokens autoregressively, which makes inference slow. Recent methods instead flow match continuous latents built by Variational AutoEncoders (VAEs), but reconstruction quality drops significantly when geometry and topology are jointly encoded, and further when the latent space is compressed. We present MeshCarve, a flow matching method that generates entirely in compact latent spaces, generating vertex positions and edge connections separately and sidestepping the difficulty of a joint compact latent. To shorten the token sequence, we propose a hierarchical sparse transformer backbone, instantiated as VertexVAE and EdgeVAE. Instead of encoding fields over the surface voxels, both VAEs anchor on discrete vertices in their latent spaces, which drastically reduces the token sequence length, and our spatial-aware compression shortens it further without costing reconstruction. VertexVAE directly encodes vertex occupancy. For connectivity, we propose vertex-link encoding, which turns arbitrary connectivity between vertices into fixed-length continuous per-vertex embeddings and recovers complex artistic topology faithfully. MeshCarve combines these VAEs with an anchor generator and flow matches on the shortened token sequences. It shows advantages over state-of-the-art autoregressive and flow matching methods on Objaverse and generalizes to Toys4K. To the best of our knowledge, it is among the first artisan mesh generation methods whose every generative stage runs in a spatially compressed latent, with a token sequence only a fraction of the most compressed previous autoregressive and flow matching works.
Xiyu Wang, Ruocheng Wu, Yufei Wang +3
Nanyang Technological University · SparAI Inc. · University of Texas at Austin
We present MeshFlow, a new method for generating artist-like 3D meshes. Current mesh generators often adopt Auto-Regressive (AR) next-token prediction, a natural choice given the discrete nature of mesh topology. However, AR methods scale poorly because the inference cost is quadratic in mesh size. They also require discretizing the vertex coordinates, which introduces quantization errors. To address these challenges, we introduce a Variational Autoencoder (VAE) that, supervised with a contrastive loss, represents both continuous vertex positions and discrete connectivity in a continuous latent space. This latent space is significantly more compact than prior token-based mesh representations. We then build a 3D generator based on a Rectified Flow transformer, generating all mesh vertices and edges in parallel. Our model generates meshes 18x faster than the fastest AR generator while also achieving excellent accuracy across standard mesh-generation metrics. Homepage: https://mesh-flow.github.io/, Code: https://github.com/facebookresearch/meshflow
Weiyu Li, Antoine Toisoul, Tom Monnier +4
Meta AI · Hong Kong University of Science and Technology