TripoSplat

From the project README

Single image → 3D using TripoSplat (Tripo AI / VAST AI Research), wired into Modly as a standard image-to-mesh model.

TripoSplat is a feed-forward Gaussian Splatting model: one forward pass turns an image into up to 262k 3D Gaussians — fast, no long diffusion-step blowup.

Modly is mesh-centric (the viewer and exporter only handle .glb meshes), so this extension reconstructs the Gaussians into a watertight, vertex-colored .glb:

The mesh path is ported from ComfyUI's native SplatToMesh (splatmesh.py): each Gaussian is rasterized as its oriented 3-sigma covariance disk into a density grid, the iso-surface is extracted with Surface Nets at an Otsu-picked level, and vertex colors are sampled from a co-splatted color volume. This respects each splat's opacity and scale+rotation, so the surface fills solidly instead of pitting. A mesh is still a lossy approximation of the splats; expect a clean, recognizable shape rather than the fidelity of a native splat render.

The mesh keeps the Gaussians' color (SH DC → vertex colors sampled from the color volume).