LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow

Hang Long1,2,*, Tianhao Zhao1,2,*, Junkai Lin1,2, Youjia Zhang1,2, Huipeng Guo1, Rendong Liang2, Jiale Xu2, Jozef Hladký3, Matthias Nießner4, Yuanming Hu2, Wei Yang1,†

1Huazhong University of Science and Technology · 2Meshy AI 3Independent Researcher · 4Technical University of Munich

* Equal contribution  ·  † Corresponding author This work was completed during internships at Meshy AI.

arXiv Code Weights Results
V-Flow → T-Flow Factorized generation
Controllable N Vertex budget
≈ 5s + 5s Two-stage sampling
LATO.2 vertex flow, topology flow, mesh generation, multi-part generation, and topology-adaptive editing overview
LATO.2 overview. V-Flow generates vertex positions under a controllable vertex count, while T-Flow predicts connectivity from the realized vertices. The factorization supports high-quality generation, scalable part-wise synthesis, and topology-adaptive editing.

Abstract

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 shared 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 realized vertices, with both stages anchored to a shared coarse voxel scaffold. Dedicated VAEs recover vertices at sub-voxel precision and embed discrete connectivity into a continuous latent space.

This factorization uniquely enables part-wise high-resolution generation and topology-adaptive editing. Experiments show that LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality.

Core Idea

Separate what the mesh is from how it connects

Vertex positions are continuous geometry. Connectivity is a relational structure over vertices that already exist. LATO.2 gives each signal its own latent space and flow.

Stage 01 · Geometry

Vertex Flow

Generate a structured vertex latent on a shared voxel scaffold, then decode a controllable number of high-resolution vertices with learned sub-voxel offsets.

Stage 02 · Connectivity

Topology Flow

Condition on the realized vertices, sample per-vertex topology features, and decode symmetric pairwise edge probabilities before recovering faces.

p(V, E | c) = p(V | c) · p(E | V, c)

Pipeline

Factorized Mesh Generation

V-VAE recovers precise geometry with sub-voxel offsets, T-VAE maps discrete connectivity into a continuous latent space, and V-Flow and T-Flow generate the two factors in sequence.

LATO.2 V-VAE, T-VAE, V-Flow and T-Flow pipeline
Overview of LATO.2 pipeline. LATO.2 factorizes explicit mesh generation into vertex and topology synthesis. V-VAE learns sparse vertex latents and reconstructs high-resolution vertex positions with sub-voxel offsets, while T-VAE learns topology latents through vertex-aware attention. At inference, V-Flow generates vertices conditioned on images with controllable vertex counts, and T-Flow predicts connectivity conditioned on the generated vertices.
01

V-VAE Sub-voxel offsets

Progressive sparse decoding predicts an offset from each surviving voxel center, recovering precise vertex positions while reducing quantization error.

02

T-VAE Continuous topology latent

T-VAE encodes the discrete adjacency relations into continuous per-vertex topology latents that can be decoded back into pairwise connectivity.

03

V-Flow Generate vertices

The target vertex count is injected into V-Flow together with visual features, guiding the vertex latent that V-VAE decodes into the realized vertex set.

04

T-Flow Generate connectivity

Conditioned on the realized vertices, T-Flow samples the topology latent; T-VAE then decodes pairwise connectivity and face loops recover the mesh.

Generation Result

One-pass Mesh Generation

V-Flow first generates the vertex set, and T-Flow then predicts connectivity conditioned on the realized vertices to form the final mesh.

Set 01 / 01
Faces —
Vertices —
Faces —
Vertices —
Faces —
Vertices —

Controllable Mesh Complexity

By injecting different target vertex counts into V-Flow under the same generation condition, LATO.2 produces meshes with controllable mesh complexity.

Page 01 / 01

Actual vertices: —

Actual vertices: —

Actual vertices: —

Actual vertices: —

Actual vertices: —

Actual vertices: —

Application 01

Part-wise Generation & Refinement

Each part uses the full latent capacity before T-Flow reconnects the recomposed vertex set, enabling substantially higher mesh resolution.

LATO.2 part-wise generation pipeline with controllable structure planning and independent part generation
Part-wise generation pipeline. A controllable planner introduces part-level boxes, each part uses the full latent capacity for vertex generation, and T-Flow predicts coherent connectivity over the recomposed vertex set.
How it works

Resolution scales with parts

The coarse structure is partitioned and each part is normalized to occupy the complete latent volume. V-Flow generates every part at full capacity, then the vertices are transformed back and connected by T-Flow. The result is denser and more detailed than monolithic generation at the same latent resolution.

Result 01 / 01
0%
Faces —
Vertices —
Parts —

Application 02

Topology-adaptive Mesh Editing

Edited or recomposed vertices are passed back to T-Flow so connectivity can adapt without hand-authored topology operations.

How it works

Edit vertices, regenerate connectivity

Users can stitch vertex sets from different meshes or rotate and translate a local part. T-Flow regenerates connectivity around the affected regions, bridging new junctions while avoiding stretched faces and self-intersections. No mesh-specific re-optimization is required.

Editing -
Result 01 / 01

Original

Original Vertices

Edited Vertices

Edited

Comparison

Stronger geometry and cleaner connectivity

On geometry-conditioned mesh generation, LATO.2 achieves the strongest performance across all reported surface metrics.

Qualitative comparison of LATO.2 against autoregressive and flow matching mesh generation methods
Qualitative comparison. Autoregressive and previous flow-based methods may introduce missing regions, broken surfaces, or coarse connectivity. LATO.2 preserves fine structures while producing more complete, coherent meshes.

Reference

Citation

@misc{long2026lato2factorized3dmesh,
      title={LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow}, 
      author={Hang Long and Tianhao Zhao and Junkai Lin and Youjia Zhang and Huipeng Guo and Rendong Liang and Jiale Xu and Jozef Hladký and Matthias Nießner and Yuanming Hu and Wei Yang},
      year={2026},
      eprint={2607.10623},
      archivePrefix={arXiv},
      primaryClass={cs.GR},
      url={https://arxiv.org/abs/2607.10623}, 
}