ComfyUI-LATO2 turns TRELLIS.2/Hunyuan3D meshes into usable low-poly, 990k to 3k faces

sakalond · reddit · 2026-11-03

Reddit user sakalond released ComfyUI-LATO2, a node wrapping LATO.2 (HUST & Meshy, MIT license), a model that regenerates a fresh low-poly mesh (200–5000 vertices) from the dense 500k–1M-triangle outputs of image-to-3D models like TRELLIS.2 and Hunyuan3D — shown taking a well from 990k to 3k faces.

Not a decimator: it places vertices and predicts edges, so topology mostly follows the form.
The node rebuilds LATO.2's raw output (overlapping faces, holes, flipped faces) into a closed outward-facing surface, strips TRELLIS.2's hidden inner wall via voxel flood fill, and preserves position/scale for drop-in replacement.
Ships pure PyTorch/NumPy stand-ins for CUDA-only deps (ovoxel, flash-attn) so it runs on Windows; verified identical voxels over millions of samples.
Needs an NVIDIA GPU with 8GB VRAM; 30–60s per run on a 3090/4080. Four example workflows included.

Caveats: thin parts may fail (try 2–3 seeds); geometry only, no textures/UVs. GitHub: github.com/sakalond/ComfyUI-LATO2.

Original post →

ComfyUI-LATO2 turns TRELLIS.2/Hunyuan3D meshes into usable low-poly, 990k to 3k faces

More from Multimodal