Hang Long, Tianhao Zhao, Junkai Lin, Youjia Zhang, Huipeng Guo, Rendong Liang, Jiale Xu, Jozef Hladký, Matthias Nießner, Wei Yang
6 min
Existing 3D mesh generation models often struggle with the hybrid nature of meshes, where continuous vertex coordinates must be reconciled with discrete, combinatorial connectivity. Previous approaches typically encode both in a single latent space, leading to entangled representations that manifest as drifting vertices and broken surfaces. LATO.2 investigates whether decomposing this process into two distinct, sequential stages—vertex generation followed by connectivity prediction—can produce higher-fidelity, structurally coherent meshes.
LATO.2 introduces a factorized flow matching framework. First, a Vertex VAE (V-VAE) compresses vertex positions into a structured latent, using a per-vertex drift vector to refine positions at sub-voxel precision. Second, a Topology VAE (T-VAE) learns to represent connectivity as per-vertex features, conditioned on the realized vertex positions.
During generation, the pipeline proceeds in two stages:
This factorization allows for "part-wise" generation, where the mesh is partitioned into smaller sections to achieve higher resolution, and "topology-adaptive editing," where modifying vertex positions automatically updates the connectivity without requiring re-optimization of the entire mesh.
By separating geometry from topology, LATO.2 addresses the fundamental difficulty of modeling discrete combinatorial structures within continuous generative spaces. This approach not only improves the geometric fidelity and connectivity quality of generated meshes compared to state-of-the-art methods but also provides practitioners with greater control over mesh resolution and structure. The ability to perform topology-adaptive editing makes this framework particularly useful for production pipelines where artist-like, well-structured topology is required for rigging and animation.
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.
Sam: Precisely. And separating those steps unlocks something useful for artists, too. The paper describes a capability they call "topology-adaptive editing." If a designer stretches a character's arm to make it longer, the system doesn't just warp the existing surface and break it. Instead, it automatically recalculates the best way to connect the points for the new shape, keeping the mesh clean and ready for animation. That's the kind of thing that matters enormously in a real production pipeline.
Alex: So it's not just about whether the final image looks right — it's about whether the model is actually usable downstream.
Sam: That's the goal. Now, the way each stage works internally is worth understanding. The vertex stage uses something called a Variational Autoencoder — or VAE. Think of it as a smart compression tool. It takes complex 3D data and squashes it down into a compact summary, then learns how to reconstruct a high-resolution shape from that summary.
Alex: Like a zip file for 3D models?
Sam: A smart zip file, yes. The key detail is that the Vertex VAE places points on a regular 3D grid — like graph paper in three dimensions — and then uses a small mathematical adjustment to nudge each point to exactly the right spot. That way, you get precision without needing an impossibly fine grid to begin with.
Alex: And the second stage — the topology side — how does that work?
Sam: The Topology VAE uses a mechanism called "attention," which is the same idea behind how modern language models read text. It looks at each point and essentially asks, "Who are my neighbors? Which nearby points should I be connected to?" It only draws connections that make geometric sense, so instead of a tangled mess, you get a clean, well-organized surface.
Alex: So it's a two-part assembly line. Place the dots, then connect them intelligently.
Sam: That's the core insight. And because those two tasks are handled separately, the system avoids the interference errors that plague most other approaches. The paper reports this leads to significantly more accurate geometry, particularly for complex shapes.
Alex: That raises an obvious question, though. If the stages are sequential, what happens when the first stage makes a mistake?
Sam: That's the main trade-off. Because topology generation relies entirely on the output of the vertex stage, any errors made during point placement are locked in. The topology stage assumes the vertices it receives are correct — it has no way to flag a problem and ask the geometry stage to try again.
Alex: So it's a one-way street. If a point ends up in the wrong place, the surface-building stage just has to work around it.
Sam: Exactly. That's why future work is looking at iterative refinement — a closed-loop version where the topology stage could send feedback back to the vertex stage. That would let the system catch and correct geometry mistakes before they get baked into the final mesh. It's a meaningful next step, but it's not in this version.
Alex: Are there other constraints, beyond that sequencing issue?
Sam: There's a scaling challenge. The system models how every point relates to its neighbors, and that comparison process grows in complexity as you add more points. For the mesh sizes this paper targets, it's manageable. But for extremely high-detail models, the system may eventually need a more efficient way to decide which points are worth comparing at all.
Alex: And what about the surface appearance — colors, textures, materials?
Sam: That's the remaining gap. This research focuses entirely on the skeleton of the 3D object — the geometry and the connections. Real production assets also need textures and material properties, and the current model doesn't handle those. The authors are clear that this is a foundation, not a finished pipeline.
Alex: So they've built a solid structural framework, but the "skin" of the model is still future work.
Sam: That's a fair summary. By treating the shape and the surface connections as two separate, learnable problems, LATO.2 provides a more controlled and reliable path toward 3D generation that's actually useful in practice. It's a meaningful step — not the end of the road, but a clearer direction forward. Thanks for listening to ResearchPod.