ComfyUI-LiTo

[!WARNING] Warning, uses experimental package comfy-env to attempt a one click isolated install. Will download and use pixi package manager.

https://github.com/user-attachments/assets/c3ff490b-e0d4-47cc-953b-a2236673e8bf

ComfyUI-LiTo

ComfyUI wrapper for LiTo (Surface Light Field Tokenization), Apple Research's image-to-3D Gaussian Splat generator from ICLR 2026.

Single RGBA image → 3D Gaussians (~524K splats) capturing geometry + view-dependent appearance (specular highlights, Fresnel reflections). ~4.7s on H100 after compilation.

Nodes

| Node | Description | |------|-------------| | (Down)Load LiTo Model | Downloads lito_dit_rgba.ckpt (~3GB) from Apple CDN to ComfyUI/models/lito/ | | LiTo Preprocess Image | Background removal (rembg) + crop/pad to 518×518 RGBA | | LiTo Image to 3D | DiT flow-matching sampling + Gaussian decoding | | LiTo Export PLY | Save Gaussians to standard PLY | | LiTo Preview Point Cloud | Browser-based 3D preview |

Workflow

LoadImage → LiToPreprocess → LiToImageTo3D → LiToExportPLY → LiToPreviewPointCloud
                                  ↑
                            LiToLoadModel

Installation

This pack uses comfy-env for process isolation — LiTo's heavy dependencies (pytorch3d, gsplat, nvdiffrast, flash_attn, spconv, xformers) run in their own subprocess environment, separate from ComfyUI's host Python.

CUDA wheels resolved automatically from cuda-wheels.

License

The wrapper code follows Apple's LICENSE. The model weights are governed by LICENSE_MODEL (Apple ML Research model license — non-commercial research use).

Credits

• LiTo by Jen-Hao Rick Chang, Xiaoming Zhao, Dorian Chan, Oncel Tuzel (Apple, ICLR 2026)