visualbruno/ComfyUI-Trellis2

ComfyUI Wrapper for Microsoft Trellis.2 - Native and Compact Structured Latents for 3D Generation

📦 What is ComfyUI‑Trellis2?

A community‑maintained custom node pack that brings Microsoft’s TRELLIS.2 3‑D generation pipeline into ComfyUI, the visual node‑based UI for diffusion models. It lets you build end‑to‑end workflows that turn one or many images into textured 3‑D meshes, multi‑view renders, and voxel‑based reconstructions – all inside the familiar ComfyUI canvas.

✨ Key capabilities (as listed in the changelog)

  • Multi‑view generation – nodes such as Sparse MultiView Generator, ImageCond MultiView Generator, Shape Cascade MultiView Generator, etc., create consistent geometry from several viewpoints.
  • Mesh texturing & refinement – Mesh Texturing Multi‑View, Mesh Texturing Pixal3D MultiView, and Mesh With Voxel Multi‑View Generator add UV‑mapped textures, support Pixal3D‑T models, and allow voxel‑based detail.
  • Geometry processing – nodes for remeshing, smoothing (PyMeshlab, Trimesh), hole filling (Meshlib, CuMesh), vertex welding, normal smoothing, and simplification.
  • Rendering – Render MultiView (Nvdiffrast) and Preview 3D let you visualise the result directly in ComfyUI.
  • Advanced samplers – experimental Heun, RK4, RK5 samplers for diffusion steps.
  • Support for external models – requires Facebook’s DINO‑v3 vision model and optionally TencentARC/Pixal3D‑T (needs the natten package).
  • Platform‑specific wheels – pre‑compiled binaries for CuMesh, nvdiffrast, nvdiffrec_render, FlexGEMM, and o_voxel for Torch 2.7/2.8 on Windows.

🛠️ Installation (quick‑start)

  1. Prerequisites
    • Install ComfyUI (standard or portable).
    • Download the DINO‑v3 model from HuggingFace and place it under ComfyUI/models/facebook/dinov3-vitl16-pretrain-lvd1689m.
    • (Optional) For Pixal3D‑T support, install the natten package from its GitHub repo.
  2. Install the provided wheels – choose the folder matching your Torch version (2.7.0 or 2.8.0) and run the pip install …/*.whl commands shown in the README. There are separate commands for a normal Python install and for the ComfyUI portable bundle.
  3. Install Python dependencies
    pip install -r ComfyUI/custom_nodes/ComfyUI-Trellis2/requirements.txt
  4. Custom builds (optional) – if you want the author’s patched versions of o_voxel, CuMesh, or FlexGEMM, clone the linked repos and install them.
  5. Restart ComfyUI – the new Trellis2 nodes will appear under a Trellis2 category.

🚀 How to use it

  1. Add a “PreProcess Image” node (optionally with remove_background via rembg).
  2. Feed the image(s) into a Multi‑View Generator – pick the style you need (Sparse, ImageCond, Shape, etc.).
  3. Optionally run a “Mesh With Voxel … Generator” to create a voxel representation, then convert to mesh with the Voxel to Mesh node.
  4. Refine geometry using Smooth Mesh, Fill Holes With Meshlib, Weld Vertices, or Simplify Trimesh.
  5. Texture the mesh with Mesh Texturing Multi‑View (or the Pixal3D‑specific node for higher‑quality texture maps).
  6. Render the result with Render MultiView (Nvdiffrast) or view it live via Preview 3D.
  7. Export using the Export Mesh node – the output is compatible with the preview node and can be saved as standard 3‑D formats.

The repository includes an example_workflows folder with ready‑made ComfyUI graphs for common tasks such as Projection_Blender_Qwen_XViews, Projection MultiView Texturing, and Pixal3D pipelines.

📚 Where to look for help

  • Discord community (linked in the acknowledgements) – active users share workflow tips and bug fixes.
  • “Blackwell Fix” – a community patch for a known issue, referenced in the README.
  • The author’s forks of Trellis.2, CuMesh, and FlexGEMM contain additional documentation and the latest bug‑fixes.

🔗 Related projects

  • Microsoft/TRELLIS.2 – the original research code that performs multi‑view diffusion‑based 3‑D generation.
  • ComfyUI – the node‑based UI that this wrapper plugs into.
  • Pixal3D – a high‑quality 3‑D model dataset used by some of the nodes.

ComfyUI‑Trellis2 is essentially a bridge that makes Microsoft’s cutting‑edge 3‑D diffusion pipeline accessible to artists and developers who prefer a visual workflow. It bundles the necessary compiled extensions, provides a rich set of geometry‑processing nodes, and ships example graphs to get you generating textured meshes in minutes.

Related

  • Project
    PozzettiAndrea/ComfyUI-TRELLIS2

    Custom ComfyUI nodes that implement Microsoft's TRELLIS.2 model to generate high-quality 3D meshes with PBR materials from a single image.

  • Project
    Comfy-Org/ComfyUI_frontend

    The official front-end for ComfyUI, providing a visual node-based interface and a robust extension API for designing AI workflows.

  • Project
    YanWenKun/ComfyUI-Windows-Portable

    A ready‑to‑run Windows package that bundles ComfyUI (a node‑based UI for Stable Diffusion) with 40+ custom nodes and 300+ pre‑compiled Python packages, letting users launch AI image/video generation instantly on an NVIDIA GPU.

  • Project
    WASasquatch/was-node-suite-comfyui

    WAS Node Suite v3 is a MIT‑licensed collection of 481 custom nodes for ComfyUI, adding extensive image‑processing, text, animation, and three‑js capabilities. Install via the ComfyUI Manager or manual clone, then enable optional features in `config.yaml`. No runtime package installs; everything ships in‑repo, with security‑first defaults.

  • Project
    Comfy-Org/comfy-cli

    comfy‑cli is a Python command‑line tool that installs, launches, and manages ComfyUI (an open‑source generative‑media engine). It handles environment setup, custom‑node installation, model downloads, workflow execution, and can offload jobs to the hosted Comfy Cloud service. Features include fast `uv`‑based dependency resolution, JSON‑structured output for LLM agents, shell completion, and utilities for testing PRs and managing background servers.