meshgpt-pytorch

Why we included this project

Researchers and graphics engineers wanting to try transformer-based mesh generation will find this the most usable reference implementation of the MeshGPT architecture. It follows the paper's two-stage design: a mesh autoencoder that tokenizes triangle faces into discrete codes via residual vector quantization, then a decoder-only transformer that generates new geometry from those face-token sequences. The repo handles variable-length meshes, includes a Hugging Face accelerate trainer wrapper, and has an optional text-conditioning path with classifier-free guidance if you want to experiment with text-to-3D pipelines. There are also community-trained checkpoints on Hugging Face, so you can load a working model and evaluate generation quality without training from scratch.