# I made a ComfyUI node for LATO.2: turn dense TRELLIS.2 / Hunyuan3D meshes into usable low-poly meshes — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/c7a39fd9d8572269

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-11-03 01:29 ingested 2026-11-04 02:41

[original source ↗](https://www.reddit.com/r/comfyui/comments/1xmfrrc/i_made_a_comfyui_node_for_lato2_turn_dense/)

## retained text



```
Image-to-3D models like TRELLIS.2 and Hunyuan3D give you great shapes, but as 500k–1M triangles of soup that you can't use in a game or animate. **LATO.2** (HUST & Meshy, MIT license) is a new model that generates a *new* low-poly mesh for a shape: you ask for a vertex count (200–5000), it places the vertices and predicts how they connect. It's not a decimator, so the edges mostly follow the form. It's not hand-modelled quality either: think fairly clean, not artist-clean.

**ComfyUI-LATO2** wraps it as a ComfyUI node and makes it usable out of the box:

* **Works with ComfyUI's native 3D nodes**: plug in a `MESH` straight from native TRELLIS.2 or Hunyuan3D, or a 3D file from Load 3D; get a `MESH` for Save 3D Model and a 3D file for Preview 3D. Or just pick/upload a .glb/.obj.
* **Closed surface**: LATO.2's raw output stacks overlapping faces, leaves holes and flips faces at random (last image). The node rebuilds the faces from LATO.2's edge predictions into one closed surface with the faces pointing outward.
* **Inner walls removed first**: TRELLIS.2 meshes have a hidden second copy of their surface inside; a voxel flood fill strips it so LATO.2 doesn't mesh both walls.
* **Same position and scale** as the input, so it drops in as a replacement.
* **Windows and Linux, nothing to compile**: LATO.2 normally needs CUDA extensions (o\_voxel, flash-attn, ...) that have no Windows builds. The node ships pure PyTorch/NumPy stand-ins (the voxelizer gives identical voxels to the original, checked on millions of voxels), and a one-command installer sets up its own environment so your ComfyUI Python stays untouched.

**Install**: `git clone` [`https://github.com/sakalond/ComfyUI-LATO2`](https://github.com/sakalond/ComfyUI-LATO2) into `ComfyUI/custom_nodes`, then run `python install_lato2.py` once in the node's folder (about 12 GB: environment + weights) and restart. It's also submitted to the Comfy Registry as `comfyui-lato2`, but that listing is still pending review, so ComfyUI Manager may not find it yet. The *LATO.2 Setup Check* node tells you if anything is missing. Four example workflows are included (Browse Templates → Extensions → ComfyUI-LATO2), including one that plugs into ComfyUI's own TRELLIS.2 template.

**Needs**: NVIDIA GPU with 8 GB VRAM (peaks around 7 GB). A run takes 30–60 s on a 3090/4080.

**Caveats**:

* Thin parts like windmill blades or slats can come out badly, so try 2–3 seeds.
* Geometry only: no textures or UVs, so texture it afterwards.

GitHub: [https://github.com/sakalond/ComfyUI-LATO2](https://github.com/sakalond/ComfyUI-LATO2)

All credit for the model to the LATO.2 authors: [https://github.com/LoHhhha/LATO.2](https://github.com/LoHhhha/LATO.2) (paper: [https://arxiv.org/abs/2607.10623](https://arxiv.org/abs/2607.10623)). The meshes in the images were generated with TRELLIS.2 in StableGen, my Blender add-on for AI texturing and 3D generation ([https://github.com/sakalond/StableGen](https://github.com/sakalond/StableGen)).

Happy to answer questions or take feature requests.
```



document fetched 2026-11-04 00:41 · canonical https://reddit.com/r/comfyui/comments/1xmfrrc/i_made_a_comfyui_node_for_lato2_turn_dense
