MeshAnythingV2

Why we included this project

This is the official reference implementation of MeshAnything V2, an ICCV 2025 research project that turns a rough 3D shape into a clean, artist-quality polygon mesh of the sort used in games, film, and virtual production. Instead of returning dense or noisy geometry, the model is trained on meshes built by human artists and uses an autoregressive transformer to emit faces one at a time, much as a language model writes tokens. The core idea is a tokenization scheme that represents each face with a single vertex where possible, roughly halving the sequence length and letting the model reach up to 1600 faces without extra compute. Teams building 3D content pipelines that need controllable, low-polygon output they can edit or rig will find the code, pretrained weights, and a hosted demo handy for experimenting with automatic asset generation.