Meshy T2 cuts image-to-3D mesh generation to 6 seconds

A median of 6 seconds for end-to-end image-to-3D mesh generation is the headline number in the new Meshy T2 paper, and the method behind it is where the story actually is. The mainstream approach for native mesh generation, as the paper frames it, serializes a mesh into a token sequence and decodes it autoregressively, 'which is slow at inference and sensitive to error accumulation, making them impractical for interactive asset creation.' Meshy T2 walks away from that structure.

In its place is flow matching, run as a coarse-to-fine cascade of two models. An image-conditioned voxel flow first sketches the overall shape as a coarse occupancy scaffold. A mesh flow then populates that scaffold with per-vertex latent tokens, conditioned on the image, the scaffold, and a requested vertex budget. Underneath, a vertex-set mesh VAE encodes each vertex as one continuous latent token and decodes vertices, edge connectivity, and face winding order in a single pass, which the authors say preserves 'high-precision geometry and artist-authored topology without vertex quantization or welding.' The reported result is a median of 6 seconds and 'over an order of magnitude faster than autoregressive baselines,' with what they call 'state-of-the-art geometric fidelity.'

Why this matters if you are not in graphics research: artist-quality mesh topology is one of the last real bottlenecks in 3D asset pipelines. Tools that generate visually plausible surfaces have existed for a while, but ones that produce clean, engine-ready topology fast enough for interactive iteration have not. If a native mesh model can produce that in seconds with a controllable vertex budget and native multi-part output, the loop between an art director's revision and the next asset shrinks from an overnight batch to something closer to a design conversation.

The honest caveat is what the abstract does not say. The over-an-order-of-magnitude speed claim does not name the specific autoregressive baselines Meshy T2 was measured against, does not specify hardware, and does not sit next to concrete fidelity numbers. There is nothing here about a public checkpoint, code release, or licensing either, so real production trials and third-party verification are still open. Take the specifics as reported, not settled.

For game and film teams the direction is the part worth watching. A native mesh model built for parallel flow-based synthesis, with a vertex budget an artist can dial and multi-part components that emerge from the generation itself, looks like the shape of the next asset pipeline, and any vendor still built around autoregressive mesh tokenization now has a roadmap question to answer.