MeshGPT is an innovative approach to generating triangle meshes that emulates the compactness found in artist-created meshes, unlike the dense meshes obtained from neural fields. Taking inspiration from advancements in large language models, MeshGPT employs a sequence-based approach to generate triangle meshes by encoding local mesh geometry and topology into latent quantized embeddings, which are then decoded into triangles. This approach allows for the effective reconstruction of the mesh. The model, trained on this learned vocabulary, can generate new triangle meshes that closely resemble the efficient triangulation patterns seen in human-crafted meshes. Notably, MeshGPT outperforms existing mesh generation methods with higher shape coverage and improved FID scores.
https://nihalsid.github.io/mesh-gpt/