The dominant paradigm for high-fidelity 3D generation relies on a VAE-Diffusion pipeline, where the VAE's reconstruction capability sets a firm upper bound on generation quality. A fundamental challenge limiting existing VAEs is the representation mismatch between ground-truth meshes and network predictions: GT meshes have arbitrary, variable topology, while VAEs typically predict fixed-structure implicit fields (\eg, SDF on regular grids). This inherent misalignment prevents establishing explicit mesh-level correspondences, forcing prior work to rely on indirect supervision signals such as SDF or rendering losses. Consequently, fine geometric details, particularly sharp features, are poorly preserved during reconstruction. To address this, we introduce TopoMesh, a sparse voxel-based VAE that unifies both GT and predicted meshes under a shared Dual Marching Cubes (DMC) topological framework. Specifically, we convert arbitrary input meshes into DMC-compliant representations via a remeshing algorithm that preserves sharp edges using an L∞ distance metric. Our decoder outputs meshes in the same DMC format, ensuring that both predicted and target meshes share identical topological structures. This establishes explicit correspondences at the vertex and face level, allowing us to derive explicit mesh-level supervision signals for topology, vertex positions, and face orientations with clear gradients. Our sparse VAE architecture employs this unified framework and is trained with Teacher Forcing and progressive resolution training for stable and efficient convergence. Extensive experiments demonstrate that TopoMesh significantly outperforms existing VAEs in reconstruction fidelity, achieving superior preservation of sharp features and geometric details.
Figures & tables
Figure 2: TopoMesh comprises two core modules. Topo-Remesh converts imperfect, real-world meshes into clean, DMC-compliant representations while preserving sharp features. Topo-VAE takes vertices with normals as input and reconstructs the mesh in the same DMC format. With topological unification, we apply direct supervision on mesh attributes: topology, vertex positions, and face orientations.
Figure 3: (a) Compression of the sparse attention map. (b) A single, shared query aggregates all keys from the same voxel to form the output feature of this voxel.
Figure 4: Illustration of explicit mesh-level loss.
Figure 5: (a) Angle-preserving of L ∞ metric. (b) 2D illustration of L ∞ distance. P lies on the isosurface that preserves angles.
Figure 6: Remesh pipeline and execution time.
Figure 7: Visual comparison of remeshing. Our method produces clean meshes with sharp features. ∗ means a re-implementation version.
Method
Device
Thingi10K [ 65 ]
Objaverse [ 9 ]
Efficiency
CD ↓
F1 ↑
ANC ↑
CD ↓
F1 ↑
ANC ↑
Time ↓
Size ↓
Mesh2SDF [ 48 ]
CPU
1.980
0.932
0.955
3.349
0.931
0.829
552.3
66.5
ManifoldPlus [ 15 ]
CPU
1.347
0.976
0.981
0.620
0.991
0.780
79.4
92.7
Dora [ 2 ]
GPU
1.492
0.972
0.970
1.057
0.987
0.961
116.3
112.6
Sparc3D [ 23 ]
GPU
1.436
0.978
0.975
2.864
0.970
0.929
175.9
121.6
Ours
GPU
1.479
0.978
0.984
0.964
0.988
0.964
18.5
28.7
Table 1: Quantitative result of remeshing. Our method preserves fine details with the highest efficiency. CD is multiplied by 105 .
Figure 8: Visual comparison of VAE reconstruction. Our method better preserves sharp features and fine details.
Method
#Latent
#Dim
Topo-Bench
Dora-Bench [ 2 ]
CD ↓
F1 ↑
F1-S ↑
ANC ↑
CD ↓
F1 ↑
F1-S ↑
ANC ↑
TripoSG [ 22 ]
4096
64
2.658
0.893
0.715
0.965
1.697
0.959
0.717
0.976
Dora [ 2 ]
4096
64
2.167
0.905
0.754
0.968
1.814
0.964
0.768
0.977
Hunyuan3D-2.1 [ 45 ]
4096
64
2.538
0.888
0.767
0.965
1.606
0.954
0.770
0.980
Trellis [ 55 ]
12832
8
18.616
0.583
0.308
0.893
14.716
0.715
0.279
0.915
Direct3D-S2 [ 54 ]
76386
16
2.713
0.813
0.694
0.920
2.313
0.881
0.819
0.968
Table 2: Quantitative result of VAE. All methods except Trellis produce meshes at 10243 resolution. Our method outperforms other methods in preserving sharp features. CD is multiplied by 105 .
Figure 9: Dihedral angle distributions from different remeshing methods.
Figure 10: Comparison of direct mesh-level supervision versus rendering-based supervision.
Method
CD ↓
F1 ↑
F1-S ↑
ANC ↑
render loss
1.731
0.776
0.711
0.932
mesh-level loss
0.150
0.975
0.991
0.999
32res
1.812
0.933
0.790
0.968
64res
1.693
0.940
0.869
0.983
128res
1.126
0.973
0.915
0.995
Table 3: Quantitative results of ablations.
Method
Hunyuan3D-2.1 [ 45 ]
Trellis [ 55 ]
Direct3D-S2 [ 54 ]
Ours
FID ↓
59.43
59.61
45.33
42.48
KID ( ×103 ) ↓
5.97
6.03
5.47
4.63
Table 4: Quantitative generation results on Toys4K.
Figure 11: Visual Comparisons of image-to-3D generation.
Autoencoders learn compressed representations by encoding data into latent codes and decoding them for reconstruction, generation, denoising, or anomaly detection.
In this paper, we introduce LATO, a novel topology-preserving latent representation that enables scalable, flow matching-based synthesis of explicit 3D meshes. LATO represents a mesh as a Vertex Displacement Field (VDF) anchored on surface, incorporating a sparse voxel Variational Autoencoder (VAE) to compress this explicit signal into a structured, topology-aware voxel latent. To decapsulate the mesh, the VAE decoder progressively subdivides and prunes latent voxels to instantiate precise vertex locations. In the end, a dedicated connection head queries the voxel latent to predict edge connectivity between vertex pairs directly, allowing mesh topology to be recovered without isosurface extraction or heuristic meshing. For generative modeling, LATO adopts a two-stage flow matching process, first synthesizing the structure voxels and subsequently refining the voxel-wise topology features. Compared to prior isosurface/triangle-based diffusion models and autoregressive generation approaches, LATO generates meshes with complex geometry, well-formed topology while being highly efficient in inference.
Tianhao Zhao, Youjia Zhang, Hang Long +7
Huazhong University of Science and Technology · Peking University · Independent Researcher +1
Polygonal meshes are the standard surface representation of modern 3D pipelines, and generating high-quality meshes with artist-style topology is essential for film, gaming, and interactive 3D applications. Mainstream approaches serialize a mesh into a token sequence and decode it autoregressively, which is slow at inference and sensitive to error accumulation, making them impractical for interactive asset creation. We present Meshy T2, a fast native mesh generation framework built on flow matching. At its core is a vertex-set mesh VAE that encodes a mesh into one continuous latent token per vertex and decodes vertices, edge connectivity, and face winding order in a single pass, preserving high-precision geometry and artist-authored topology without vertex quantization or welding. Generation proceeds as a coarse-to-fine cascade of two flow-matching models: an image-conditioned voxel flow first sketches the overall shape as a coarse occupancy scaffold, and a mesh flow then populates the scaffold with per-vertex latent tokens, conditioned on the image, the scaffold, and a requested vertex budget. This design delivers three practical capabilities: interactive generation speed through parallel flow-based synthesis; effective face-count control through the requested vertex budget; and native support for multi-part assets, whose components emerge directly from the generated connectivity. In our experiments, Meshy T2 achieves state-of-the-art geometric fidelity and completes end-to-end image-to-mesh generation within a median of 6 seconds, over an order of magnitude faster than autoregressive baselines. Code and weights will be available at https://github.com/meshy-dev/meshy-t2.
Autoregressive mesh generation has gained attention by tokenizing meshes into sequences and training models in a language-modeling fashion. However, existing approaches suffer from two fundamental limitations: (i) low tokenization efficiency, which yields long token sequences and prevents scaling to high-poly meshes, and (ii) absence of geometry-aware guidance, as generation is conditioned only on global shape embeddings rather than local surface cues. We introduce MeshWeaver, an autoregressive framework that treats mesh generation as a surface weaving process by directly predicting the next vertex instead of independent coordinates. At its core is a multi-level sparse-voxel encoder that injects geometric context into the generative process in three complementary ways: providing voxel features as vertex representations, guiding token prediction via cross-attention to voxel features, and serving as a structural scaffold that constrains generation around the input surface. Our hierarchical design enables coarse-to-fine vertex prediction in a single decoding step, while tightly coupling the generative model with 3D geometry. Extensive experiments demonstrate that MeshWeaver achieves a state-of-the-art compression ratio of 18%, can generate meshes with up to 16K faces, and significantly improves geometric fidelity over prior approaches.