Meshy T2: fast AI 3D meshes with artist-grade topology, code soon
A flow-matching model turns one image into a poly-budgeted, multi-part mesh with no retopo — 10x faster than autoregressive gen. Open-source is imminent.
Copy markdown
Native meshes, not triangle soup
Meshy T2 generates geometry directly with flow matching: a VAE encodes each vertex as one continuous latent token, so it keeps artist-authored topology with no vertex quantization or welding. You get meshes you can rig and edit, not the dense marching-cubes output that forces a retopo pass.
Image to mesh in a median 6 seconds
A two-stage cascade — voxel flow for the shape scaffold, then mesh flow for the detail — turns a single image into a mesh in about 6 seconds, over an order of magnitude faster than autoregressive baselines. Fast enough to iterate on an asset in the loop instead of batching overnight.
You set the polygon budget
Request a vertex or face budget and the model targets it, and multi-part assets emerge straight from the generated connectivity with no separate part-wise step. For game and real-time work that is the ballgame: a low-poly prop and a hero asset from one pipeline.
The catch: a paper, not a package yet
Code and weights are not public yet — Meshy's repo says an open-source release is in prep ("stay tuned," dated Aug 3), and the speed numbers are the authors' own. Meshy already runs a commercial image-to-3D API, so watch meshy-dev/meshy-t2 if you want to self-host clean-topology 3D generation.