Adapted from Rendong Liang’s article on the Meshy blog.
When you evaluate a generated 3D asset, the preview image answers only part of the question. What does the wireframe look like? Are separate components still separate? Can the geometry fit your application’s budget?
Meshy T2 explores these questions by generating mesh geometry and connectivity directly.
Represent the mesh as a set
Autoregressive mesh generators serialize geometry into tokens and produce them sequentially. Longer sequences increase inference time and create opportunities for earlier errors to affect later predictions.
The Meshy T2 technical report describes a different representation: a vertex-set variational autoencoder, with one continuous latent token per vertex.
Its decoder reconstructs vertex positions, edge connectivity, and face winding together. Coordinates are not quantized, and coincident vertices are not automatically welded.
That distinction matters because sharing a position does not necessarily mean two vertices should become one topological element.
Separate overall shape from mesh structure
Generation uses two flow-matching stages:
- Voxel flow: use the reference image to establish a coarse occupancy scaffold.
- Mesh flow: generate vertex tokens conditioned on the image, scaffold, and requested vertex budget.
The mesh decoder then turns those tokens into geometry and connectivity.
Parallel generation here does not mean the entire process has one inference step. The flow models iteratively update the latent representation; the final mesh decoder reconstructs its outputs together.
The generated connectivity also supports disconnected components within one asset, without requiring a separate generation-and-stitching process for each part.
Control complexity during generation
The blog demonstrates budgets of 500, 1,000, 2,000, and 4,000 vertices.
This is budget-range control, rather than a guarantee of an exact triangle count. Padding tokens carry existence flags so the model can discard surplus vertices during decoding.
For developers, the useful question is how well each budget preserves the structure required by the application. A silhouette, a narrow support, and a small decorative detail may deserve different priorities.
What does “2 seconds” refer to?
The published paper reports a median end-to-end image-to-mesh time of 6 seconds.
The Meshy blog separately reports engineering-optimized generation at approximately 1 second, or approximately 2 seconds including the WebApp/API service round trip.
These figures describe different reported execution settings. The two-second figure should not be read as a universal latency guarantee for every input or deployment.
Inspect the output beyond the render
When evaluating a method like this for your own workflow, useful checks include:
- Whether important shapes survive at the intended budget.
- Whether connectivity matches the object’s structure.
- Whether the asset behaves correctly after import into your target tools.
- How much additional editing the intended use requires.
The full Meshy article includes architecture diagrams, wireframe examples, and budget comparisons.
What is your biggest obstacle when using generated 3D assets: generation time, topology, polygon count, or cleanup after import?
Top comments (0)