tianhaozhao668/lato

LATO is an AI research project for generating 3D meshes, the wire frame structures that make up 3D objects in games, films, and design software. Specifically, it focuses on producing meshes that are topology preserving, meaning the generated geometry has correctly connected surfaces rather than an approximation that needs to be reconstructed afterward. Most AI based 3D generation methods work with implicit representations, mathematical functions that describe a shape, and then convert them into a mesh using algorithms like Marching Cubes. This extra conversion step often introduces messy geometry. LATO takes a different approach: it treats the mesh directly as a Vertex Displacement Field, a way of describing where the mesh's corner points sit relative to the surface, and compresses that into a structured latent space, a compact learned representation the AI model can work with directly. The system uses a sparse transformer, an efficient type of neural network, as the encoder that compresses mesh topology, and a decoder with specialized heads to reconstruct vertices and edges. The result, according to the paper, is that it can generate clean artistic meshes in seconds without the heavy memory cost that normally makes working with dense meshes expensive. In practice, you provide a 3D object file, run an inference script, and get back a reconstructed or generated mesh. Pretrained model weights are available on Hugging Face. The project is implemented in Python and its accompanying paper was accepted at ICML 2026. It is aimed at researchers and developers working on 3D AI content generation.