explaingit

pozzettiandrea/comfyui-lito

★ 63PythonAudience · researcherComplexity · 5/5ActiveLicenseSetup · hard

TLDR

ComfyUI node pack that wraps Apple's LiTo research model so a single image becomes about 524,000 Gaussian Splats representing a 3D object with reflectance.

Mindmap

mindmap
  root((ComfyUI-LiTo))
    Inputs
      Single object image
      LiTo checkpoint
      ComfyUI workflow
    Outputs
      Gaussian Splats
      PLY file
      Browser preview
    Use Cases
      Convert image to 3D splats
      Capture shiny reflections
      Prototype Apple LiTo
    Tech Stack
      Python
      ComfyUI
      PyTorch3D
      gsplat
      CUDA

Things people build with this

USE CASE 1

Run Apple's LiTo from a ComfyUI workflow on a single object photo

USE CASE 2

Export a 3D object as a PLY of Gaussian Splats from one image

USE CASE 3

Preview splat output in the browser before saving to disk

USE CASE 4

Prototype non commercial research pipelines that need surface reflectance

Tech stack

PythonComfyUIPyTorch3DgsplatCUDA

Getting it running

Difficulty · hard Time to first run · 1day+

Needs an Nvidia GPU plus comfy-env to isolate pytorch3d, gsplat, nvdiffrast, flash_attn, spconv, and xformers from the main ComfyUI environment.

Wrapper code uses Apple's license and the LiTo model weights have a separate Apple research license restricting use to non-commercial research only.

In plain English

This repository is a small bridge between two existing pieces of AI software. The first is ComfyUI, a popular node-based interface where people chain together AI tasks by dragging boxes onto a canvas. The second is LiTo, short for Surface Light Field Tokenization, a research model published by Apple at the ICLR 2026 conference. LiTo takes a single image of an object and produces a 3D representation of it as a cloud of Gaussian Splats, a format that captures both shape and how the surface reflects light, including effects like shiny highlights and angle-dependent reflections. The wrapper makes LiTo usable from inside ComfyUI as a normal workflow. The pack adds five nodes. One downloads the LiTo model checkpoint, about three gigabytes, from Apple's content delivery network. Another preprocesses your input image by removing the background and resizing it to 518 by 518 pixels with transparency. The main node runs LiTo itself, taking the prepared image and producing roughly 524,000 splats that describe the object in 3D. The remaining two nodes save the result as a standard PLY file or show it as a point cloud preview in your browser. According to the README, a full run takes about 4.7 seconds on an Nvidia H100 graphics card after the model is compiled. Installation uses an experimental helper called comfy-env, which is also written by the same author. The reason is that LiTo depends on a long list of heavy graphics and machine-learning libraries: pytorch3d, gsplat, nvdiffrast, flash_attn, spconv, and xformers. Forcing all of those into your main ComfyUI Python environment often causes version conflicts, so comfy-env runs them in a separate subprocess with their own isolated set of packages. The matching CUDA wheel files come from a companion site the author also maintains. The wrapper code is released under Apple's license, and the model weights themselves are under a separate Apple research license that restricts use to non-commercial research. The README credits the original LiTo paper authors at Apple.

Copy-paste prompts

Prompt 1
Walk me through installing ComfyUI-LiTo using comfy-env and downloading the 3 GB Apple checkpoint
Prompt 2
Show me how to build a ComfyUI workflow that removes the background, runs LiTo, and saves a PLY
Prompt 3
Explain why ComfyUI-LiTo needs pytorch3d, gsplat, nvdiffrast, and flash_attn isolated from the main env
Prompt 4
Help me preview LiTo's 524k splat output in the browser and tune the H100 inference time
Prompt 5
Sketch a research notebook that compares LiTo splats against a diffusion based image to 3D baseline
Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.