What it does MeshAnything takes a 3D input—either a point cloud with normals or an existing mesh—and generates a new triangle mesh with up to 800 faces. It runs autoregressively, predicting faces one by one rather than dumping a dense voxel grid through Marching Cubes. The result is meant to look like something a human modeler actually built: relatively clean topology, sensible edge flow.

The interesting bit The twist is the autoregressive angle. Instead of treating mesh generation as regression or diffusion over vertices, it tokenizes faces and generates them sequentially with a transformer—language-model style, but for triangles. The authors explicitly contrast this with “feed-forward 3D generation methods,” suggesting those often produce blobs that collapse under low face budgets.

Key highlights

• Accepts .obj meshes or .npy point clouds (shape (N, 6) with coordinates + normals)
• ~7 GB VRAM, ~30 seconds per mesh on an A6000
• Preprocessing with Marching Cubes available via --mc flag
• Gradio demo and pip-installable (pip install git+https://github.com/buaacyw/MeshAnything.git)
• Weights and HuggingFace space hosted; training code lives in the newer V2 repo

Caveats

• Hard ceiling of 800 faces; sharp inputs work better, blurry ones suffer
• Input must be +Y-up and gets normalized to a unit bounding box
• The authors warn that feed-forward generated shapes often produce “bad results” due to insufficient sharpness

Verdict Worth a look if you need lightweight, artist-style meshes from scans or reconstruction outputs. Skip it if you need high-poly detail or are feeding it mushy NeRF blobs.