Source-linked AI summary
LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow
Hang Long, Tianhao Zhao, Junkai Lin, Youjia Zhang, Huipeng Guo, Rendong Liang, Jiale Xu, Jozef Hladký, Matthias Nießner, Yuanming Hu, Wei Yang
TL;DR
Existing topology-aware mesh generators entangle continuous vertex geometry with discrete connectivity, complicating flow learning and producing broken surfaces. LATO.2 factorizes generation into vertex and vertex-conditioned connectivity flows over dedicated structured latents, and experiments report stronger geometric fidelity and connectivity quality together with part-wise high-resolution generation and topology-adaptive editing.
Problem
Joint latent modeling entangles spatial vertex positions with discrete connectivity, complicating flow learning and producing broken surfaces.
Method
LATO.2 uses separate flow-matching stages and dedicated VAEs for vertex generation and vertex-conditioned topology generation over structured latents anchored to a shared voxel scaffold.
Results
LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality while supporting part-wise high-resolution generation and topology-adaptive editing.
Takeaways & Limitations
The factorization makes topology adaptive to vertex-level intervention and supports higher-resolution synthesis by partitioning the voxel scaffold.
Takeaways & Limitations
Topology generation cannot correct geometric inaccuracies from the vertex stage, and pairwise vertex relationships create quadratic decoding complexity at larger mesh resolutions.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
LATO.2 addresses the difficulty of generating artist-like mesh topology by factorizing vertex and connectivity synthesis into separate flow stages. This design supports higher-resolution part-wise generation and topology-adaptive editing while improving geometric fidelity and connectivity quality.
- Artist-created meshes combine adaptive vertex placement with coherent connectivity, but neural-field pipelines typically extract dense, irregular surfaces lacking this structure.Such structure supports rigging, deformation, shading, storage, and rendering.
- Jointly modeling spatial vertices and discrete connectivity entangles statistically different signals, complicating flow learning and producing broken surfaces.Conditioned on realized vertices, connectivity can instead be represented through continuous per-vertex features decoded into edge probabilities.
- LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality.The reported advantages also include part-wise high-resolution generation and topology-adaptive editing.
- LATO.2 decomposes explicit mesh generation into vertex and connectivity flows over structured latents anchored to a shared voxel scaffold.A per-vertex drift vector compensates finest-resolution quantization error, while topology latents decode into pairwise edge probabilities.
- The factorization enables vertex-induced part-wise high-resolution generation and topology-adaptive editing.Partitioning the scaffold lets each part use full latent capacity, while vertex manipulation induces corresponding connectivity.
2 RELATED WORK
Prior work spans 2D-prior reconstruction, autoregressive mesh sequences, and continuous explicit mesh generation. LATO.2 differs from compact latent approaches by explicitly separating high-resolution vertex generation from vertex-conditioned topology generation.
- 3D Shape Generation: Early 3D shape-generation methods use 2D diffusion priors either to optimize 3D representations or to reconstruct geometry from synthesized multi-view images.
- Autoregressive Mesh Generation: Autoregressive mesh methods generate vertices, edges, or faces as discrete sequences, but face-wise encoding can require up to 9 times the face number in coordinate tokens.
- Flow-based Explicit Mesh Generation: Continuous explicit methods avoid sequential generation through differentiable face probabilities, diffusion over triangle soups, or continuous face-level tokens.
- Flow-based Explicit Mesh Generation: MeshFlow and LATO use compact latent spaces for efficient geometry and connectivity generation, but still couple the two signals or generate them in parallel.
- Flow-based Explicit Mesh Generation: LATO.2 explicitly factorizes mesh generation into local sparse refinement for vertices and a relational topology latent conditioned on the generated vertex set.
3 METHOD
LATO.2 separates vertex and connectivity generation into two VAE-backed latent spaces and flow stages, with topology generated after vertices are realized. The pipeline uses sparse voxel structure, sub-voxel vertex offsets, per-vertex topology features, and vertex-conditioned connectivity to support editing and part-wise refinement.
- Factorized representations: LATO.2 trains separate Vertex and Topology VAEs to define distinct latent spaces for geometry and connectivity.The V-VAE and T-VAE replace a single shared latent representation.
- Vertex representation: The Vertex VAE encodes surface-derived Vertex Displacement Fields into sparse structured latents and decodes vertices through coarse-to-fine subdivision and pruning.At the finest level, local offsets recover vertex positions from voxel centers.
- Topology representation: The Topology VAE stores connectivity in per-vertex features aligned with realized vertices and reconstructs adjacency through masked encoding and unmasked decoding.The encoder observes ground-truth edges only through an adjacency attention mask, while the decoder predicts pairwise edge probabilities and recovers faces by loop detection.
- Factorized generation: Given a task condition, LATO.2 first generates sparse structure and vertices, then samples topology conditioned on the realized vertex set.The V-Flow samples vertex latents on active scaffold cells; the T-Flow uses vertex-position tokens and coarse geometric context to predict connectivity.
- Applications: Re-running the T-Flow after changing vertices propagates edits to connectivity, while partitioning the scaffold enables independent full-capacity generation for higher-resolution meshes.Part vertices are transformed back to the original coordinate frame before one or more topology-flow passes integrate connectivity.
4 EXPERIMENTS
Experiments evaluate LATO.2 across reconstruction, conditional generation, topology fidelity, and factorization, using diverse datasets and comparisons with autoregressive and latent-flow baselines. Results show strong geometric and connectivity quality, with benefits from sub-voxel offsets, shared coarse support, controllable vertex budgets, and multi-part generation.
- Evaluation scope: LATO.2 is evaluated on reconstruction, conditional generation, topology fidelity, and factorization against autoregressive and latent-flow baselines.The study also includes ablations of the vertex autoencoder, topology module, and vertex-count controllability.
- Datasets: Approximately 450K 3D assets plus 100K procedural synthetic meshes comprise the training data.The datasets include TRELLIS500K, Objaverse, Objaverse-XL, 3D-FUTURE, Toys4K, and ABO.
- Quantitative analysis: LATO.2 achieves the strongest overall performance across geometric and topology-related metrics for geometry-conditioned mesh generation.The comparison uses Chamfer Distance, Hausdorff Distance, and Normal Consistency for surface fidelity, alongside vertex-level occupancy metrics.
- Qualitative analysis: Qualitatively, LATO.2 preserves fine structures and maintains cleaner mesh connectivity than autoregressive and latent-flow baselines.Autoregressive methods often produce broken surfaces or missing thin structures, while latent-flow baselines may lose fine details or connectivity fidelity.
- Controllability and generation: Increasing the target vertex budget produces progressively denser meshes while preserving object-level geometry, enabling practical resolution control.Multi-part generation also produces finer geometric details and improved surface quality compared with monolithic generation.
- Ablation study: Removing the sub-voxel offset head degrades surface reconstruction while leaving sparse occupancy nearly unchanged.This isolates the offset head’s role in continuous vertex precision rather than sparse voxel classification.
5 CONCLUSION
LATO.2 factorizes explicit mesh generation into vertex synthesis and vertex-conditioned topology synthesis, enabling controllable generation and editing. The framework improves mesh quality while retaining clear scalability and production-oriented limitations.
- Factorized framework: LATO.2 decomposes mesh synthesis into dedicated vertex and vertex-conditioned topology stages with structured latent representations and separate flow models.The vertex branch uses sparse coarse-to-fine refinement with sub-voxel offset prediction, while the topology branch models connectivity over the generated vertex set.
- Controllable generation: The factorization enables explicit vertex-count manipulation, topology-adaptive editing, and scalable part-wise generation beyond monolithic latent capacity.These controls extend the framework beyond improving generation quality by allowing flexible vertex-level interventions and higher-resolution synthesis by parts.
- Controllable generation: Increasing the vertex budget from 0.2K to 4K produces progressively denser meshes with finer geometric details while topology adapts to the generated vertices.The same structure voxels are used across the comparison, and the underlying shape structure is preserved.
- Limitations: Topology synthesis cannot correct geometric inaccuracies from the vertex stage, and topology decoding has quadratic complexity in vertex-pair relationships.The complexity remains manageable at current resolutions but may require sparse candidate selection for substantially larger meshes.
- Limitations: The current framework generates geometry and connectivity but does not include production asset attributes such as UV coordinates, textures, or material properties.Future work also proposes tighter interaction between stages and broader complete-asset generation.
- Empirical outcome: LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality.The conclusion also positions the method as a foundation for scalable, controllable, and production-oriented explicit mesh generation.