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.
Figures & tables
Figure 2: LATO vs. Existing Paradigms. Mainstream topology-agnostic approaches utilize vecset or voxel latents decoded into implicit fields (e.g., SDF), relying on Marching Cubes for mesh extraction. Conversely, explicit mesh generation methods adopt per-face latents via autoregressive or diffusion models, but suffer from severe memory bottlenecks. LATO proposes T-Voxels latents to explicitly model topology, enabling the direct generation of artist-friendly meshes.
Figure 3: Overview of the LATO pipeline. We explicitly encode mesh topology by sampling surface points infused with relative displacement to their enclosing face vertices (Vertex Displacement Field, VDF). These dense features are aggregated and compressed via a sparse voxel VAE into a structured latent representation, termed T-Voxels . To reconstruct the mesh, the T-Voxels undergo hierarchical subdivision and learnable pruning to precisely instantiate high-resolution vertex locations. Simultaneously, a connection head predicts edge existence between vertex pairs, directly recovering the explicit mesh topology.
Method
CD(L2) ↓
CD(L1) ↓
HD ↓
∣ NC ∣↑
MeshGPT
0.042/0.050
0.061/0.069
0.070/0.096
0.824/0.831
PivotMesh
− /0.129
− /0.175
− /0.263
− /0.614
MeshCraft
0.150/0.243
0.212/0.327
0.282/0.299
0.531/0.625
Ours
0.042/0.040
0.061/0.056
0.065/0.092
0.847/0.834
Table 1: VAE reconstruction comparison. We compare reconstruction quality on dense/artistic meshes. Values are reported as dense/artistic for each metric; “ − ” denotes OOM.
Figure 4: Geometry conditioned generation comparison . We compare our method against state-of-the-art baselines. Existing methods result in either incomplete reconstructions or excessively dense and irregular topology. In contrast, our LATO generates hole-free meshes with well-formed topology suitable for downstream applications.
Figure 5: Topology comparison with implicit foundation models . We qualitatively compare against industrial-scale implicit baselines. While these foundation models possess scale and training resource advantages, their reliance on implicit field yields dense, irregular triangulation. In contrast, LATO generates artist-friendly edge flows
Figure 6: Image to 3D generation results.
Figure 7: Inference time comparison. The inference time evaluation is conducted on a single H100 GPU. LATO maintains rapid generation (3 ∼ 10s), whereas autoregressive methods exhibit prohibitive temporal scaling, often requiring minutes for high-fidelity outputs.
Figure 8: Scene-scale urban generation. We demonstrate LATO’s extensibility on 3D city data collected from website, we first synthesizes block-wise building and subsequently populates T-Voxel features. Because our sparse voxel is inherently compositional, these individual units can be seamlessly assembled into expansive, high-fidelity cityscapes.
Method
CD(L2) ↓
HD ↓
∣ NC ∣↑
MeshAnythingv2
0.108/0.066
0.227/0.137
0.696/0.766
MeshSilkSong
0.062/0.052
0.145/0.101
0.784/0.818
FastMesh
0.064/0.048
0.110/0.102
0.757/0.813
BPT
0.059/0.052
0.107/0.088
0.811/0.824
DeepMesh
0.051/0.046
0.092/0.089
0.828/0.827
Ours
0.043/0.044
0.084/0.081
0.832/0.835
Table 2: Quantitative comparison with topology-preserving generation methods on dense/artistic meshes (reported as dense/artistic).
Method
CD(L2) ↓
CD(L1) ↓
HD ↓
∣ NC ∣↑
w/o n(pk)
0.040
0.057
0.076
0.832
w/o F(p)
0.043
0.061
0.092
0.822
K=81,920
0.040
0.056
0.078
0.832
w/ norm(F(p))
0.042
0.059
0.082
0.827
w/ normτ(F(p))
0.041
0.058
0.081
0.832
Ours(Full model)
0.039
0.056
0.075
0.837
Table 3: Ablation study on VAE input. We evaluated the effects of various VAE inputs on reconstruction performance.
Figure 9: Ablation study on VDF and T-Voxel cross-attention in decoder. Left: we compare the voxel label (vertex, edge, face) input with our VDF input. The discrete label hinges the continuous learning using flow matching model. Right: Remove the T-Voxel cross-attention in decoder introduces hole and mismatches.
Figure 10: Effect of vertex number condition. As we increase the vertex number condition scaler cv , the model generates more vertices and triangles.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: Pipeline Details.
Generation Type
Method
Point-0
Point-1
Point-2
Point-3
Point-4
Autoregressive
MeshAnythingv2
(13.65, 460)
(22.87, 661)
(39.84, 1211)
(50.31, 1612)
(70.86, 2001)
MeshSilkSong
(42.23, 638)
(117.0, 1807)
(180.3, 2896)
(219.2, 3558)
(299.9, 4283)
FastMesh
(5.289, 4001)
(7.636, 6787)
(10.66, 9475)
(16.09, 11426)
(19.02, 14666)
BPT
(36.85, 623)
(81.25, 1404)
(109.4, 1995)
(164.2, 3190)
(251.8, 4029)
DeepMesh
(147.5, 3526)
(241.2, 5355)
(415.8, 9242)
(484, 12084)
(565.7, 13737)
Flow-based
Ours
( 5.470 , 265)
( 5.031 , 2702)
( 5.298 , 5262)
( 4.857 , 9803)
( 9.873 , 14809)
Appendix
Table 4: Relationship between inference time and actual output face number for different methods. Each method performs single-batch inference on an H100 GPU across 200 meshes with varying face numbers. Selected key points are displayed here for illustration. The coordinates of each point represent the inference time and the actual output face number, respectively.
Figure 12: More qualitative results on geometry conditioned generation comparison with autoregressive model.
Figure 13: Qualitative results on tolpology diversity.
Figure 14: More qualitative results of ablation study on T-Voxel cross-attention in decoder.
Figure 15: More qualitative results on effect of vertex number condition.
Figure 16: More qualitative results of ablation study on VDF.
Flow matching over carefully designed latent representations has recently emerged as a powerful paradigm for topology-aware mesh generation. Existing approaches, however, model vertices and connectivity jointly in a joint latent space, entangling continuous vertex geometry with discrete combinatorial structure; this complicates flow learning and manifests as drifting vertices and broken surfaces. We present LATO.2, a factorized flow matching framework that decomposes mesh generation into a vertex flow followed by a connectivity flow conditioned on the realized vertices, with both stages anchored to a shared coarse voxel scaffold. Dedicated VAEs underpin the two stages, recovering vertices at sub-voxel precision and embedding discrete connectivity into a continuous latent space. We demonstrate two advantages unique to this factorization: (i) part-wise generation, in which the scaffold is partitioned and each part synthesized at full latent capacity, yielding substantially higher-resolution meshes than a monolithic latent permits; and (ii) topology-adaptive editing, in which manipulating first-stage vertices induces the corresponding connectivity without re-optimization. Experiments show that LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality.
Hang Long, Tianhao Zhao, Junkai Lin +7
Huazhong University of Science and Technology · Meshy AI · Independent Researcher +1
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.
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