The method compresses meshes into compact continuous latents, then generates them in parallel through flow matching. The page emphasizes around one-second generation, 18x speedup versus autoregressive-style mesh generation, quantization-free vertex coordinates, and artist-like outputs.
MeshFlow is useful for 3D generation researchers and creative tools that need fast mesh assets rather than only implicit fields or point clouds. The project page links to CVPR, arXiv, a Gradio demo, code, model weights, and a direct denoising video.
