Drop a photo. Watch it take shape.
Open the 3D pane, drop in a picture, generate. The subject is cut out from the background automatically, a 3.3-billion-parameter diffusion shape model builds the geometry, and the finished model spins in a built-in 3D viewer with a gentle turntable idle. Every past generation stays one click away in a thumbnail history shelf.
Turn on Texture (PBR) and the same photo paints the mesh: a multiview diffusion stage generates albedo and metallic-roughness maps across six views and bakes them into a 2K texture atlas — a standard glTF PBR material that renders correctly in any engine.
The whole pipeline, ported — not proxied
Everything runs inside the same no-Python binary: the diffusion shape transformer, SDF decoding, marching-cubes surface extraction, UV unwrapping, the 2-billion-parameter multiview paint UNet, texture baking and inpainting, and a glTF writer. The texture stage is validated against the PyTorch reference at cosine 1.000 on the full denoiser step.
Surface extraction samples coarse-to-fine — 6× faster than a dense sweep at the highest resolution, with byte-identical geometry — which is why the default mesh detail is the reference's "fine" setting. One click downloads the whole stack (shape + texture in a single package), and it runs on a 16 GB Mac.
- No upload — your photos and models never leave the machine.
- Standard output — GLB with PBR materials, no proprietary format.
- Reference-faithful — per-component numerical oracles against the original.