TRELLIS: the open-source model behind Pic2Mesh

TRELLIS is an image-to-3D generative model released by Microsoft Research under the MIT license. It set a new bar for single-image 3D reconstruction quality and — because the weights are open — it's become the workhorse behind most affordable image-to-3D services, including Pic2Mesh.

Structured latents: the core idea

Earlier image-to-3D systems generated multi-view images first and stitched them into geometry, which caused melted backs and seams. TRELLIS instead works in a unified structured latent (SLAT) space: a sparse 3D grid where each active cell holds a latent feature describing local geometry and appearance. One diffusion model decides wherecells are active (the object's rough shape), a second refines what each cell contains. The same latent can be decoded into a mesh, a radiance field, or 3D Gaussians.

What it means in practice

• Clean, watertight-ish meshes with sensible topology for AI output
• Textures baked from the photo plus plausible hallucinated backs
• Runs in under a minute per object on a datacenter GPU
• MIT license: outputs are yours, commercial use allowed, no strings

TRELLIS vs. paid alternatives

Closed offerings like Tripo and Meshy add retopology, rigging, and PBR material extras on top of similar core quality. If you need a quick textured mesh from one photo, TRELLIS output is usually indistinguishable for the price of free.

Run TRELLIS on your photo now — no signup, 3 free per day. New to the tech? Start with how image-to-3D works.