ComfyUI-TRELLIS2 turns one image into a textured 3D mesh

A node wrapper around Microsoft's image to 3D model

ComfyUI-TRELLIS2 provides ComfyUI custom nodes for TRELLIS.2, which the readme identifies as Microsoft's state of the art image to 3D generation model, and the point of the wrapper is stated in one line, to generate high quality 3D meshes with physically based rendering materials from a single image. The material detail is the part that matters for a 3D pipeline. A mesh with a flat or vertex colour is a preview asset, whereas a mesh with a full PBR material set can go into a renderer, a game engine or a product visualisation pipeline, and the difference in effort downstream is large. The nodes live in a nodes directory, with the package entry point at the root of the repository, and there is a workflow examples directory plus a documentation directory holding a live test gallery, which is the most useful thing in the project for a newcomer, since it shows the node graph that produces a given result rather than describing it in prose.

Three install routes, ordered by speed and then by reliability

The installation section offers three options and describes them as being in order of speed and then reliability, which is a useful way to frame a choice. The first, and the recommended one, is the ComfyUI Manager, where you search for the package by name and install from the highest version displayed, with a fallback to the nightly build if that does not work. The second is the Manager again but installing through a git URL, pointing it at the repository address. The third is the manual route, which the readme calls the most reliable, and which is a clone into the custom nodes directory, an install of the requirements, and a run of the install script.

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-TRELLIS2.git
cd ComfyUI-TRELLIS2
pip install -r requirements.txt --upgrade
python install.py

The readme asks explicitly for problems to be reported, through either discussions or issues, and says the author is grateful for the help. That request is consistent with the shape of the project, a node wrapper whose failure modes are mostly installation and environment problems rather than logic problems.

The one click install is the experimental part

The very first thing in the readme is a warning, and it is worth reading before anything else. It says the project uses an experimental package called comfy-env to attempt a one click isolated install, and that it will download and use the pixi package manager. The same sentence appears in the package description in the manifest, which tells you the author wants the warning seen even in the metadata a package manager displays. The idea behind it is sensible, an isolated environment for a model with heavy native dependencies that would otherwise fight with the rest of a ComfyUI installation, and the fact that it downloads a package manager to do it is a reasonable price for that isolation. The risk is that an experimental path is exactly where breakage appears, so the ordering in the readme makes sense, with the fastest route first and the manual route described as the most reliable. The repository carries the files that make this work, a root file for the environment, a separate test configuration, the install script itself, and a prestartup script that ComfyUI runs when it starts.

Four dependencies, two of which are ComfyUI specific

The requirements file is short and tells you the shape of the dependency surface. There is the experimental environment package at a pinned version, a viewers package for 3D preview inside ComfyUI at a pinned version, a sparse attention package at a pinned version, and trimesh with its easy extra, which is the mesh library that handles the geometry. Pinning the three ComfyUI related packages exactly is the right call, because they are the ones that move in step with ComfyUI itself, while trimesh is a stable library that can float. The manifest also records where the models come from, with the description stating that models are downloaded automatically from Hugging Face, which is the usual pattern for a node wrapper around a published research model and means the first run needs a large download and a working connection. The publisher metadata in the manifest shows how the nodes appear inside ComfyUI, with a publisher identifier, a display name and an icon, which is how a custom node package is presented in the manager.

The test gallery and the example workflow images

The documentation is the interesting part of a small repository, and this one has a live test gallery published on a project pages site, linked from the readme with a preview image and a link to the gallery itself. For a generative 3D node, a gallery is the only honest documentation, because the output is a picture and the parameters that produced it are what a user actually wants to copy. The workflow examples in the repository are the same idea in file form, and the two images named in the readme, one for a pose and one for background removal, suggest the two workflow shapes people start with, generating an object and then removing the background from the input picture. The assets directory holds the icon and whatever else the nodes draw, and the repository is otherwise minimal, a license file, a readme, the package manifest, the requirements and the directories described above.

A single maintainer, a community Discord and no releases

The support structure is a small ecosystem rather than a process. Questions and feature requests go to GitHub discussions, and the readme points to a Comfy3D Discord server for help, updates and chat about 3D workflows in ComfyUI, which is a better place than the repository issues for a workflow question. The contributing section is one line, contributions are welcome and issues and pull requests can be submitted. The project is credited to the upstream model rather than to new work, and the credit list names TRELLIS.2 by Microsoft Research, which is the correct attribution for a wrapper. The version lives in the manifest as 0.2.19, and the repository publishes no GitHub releases, so the version to install is whatever the ComfyUI Manager shows, which is another reason the manual route with a pinned clone is the more predictable option for a production pipeline. The licence is MIT, and the last push was on 2026-08-24.

Editorial conclusion

ComfyUI-TRELLIS2 earns its place for a ComfyUI user who wants image to 3D generation inside a node graph rather than in a separate tool, since the model is exposed as nodes you can wire into a workflow, and since the meshes come with physically based rendering materials rather than a flat colour. It is the wrong choice if you want a stable, pinned installation, because the headline warning says the project uses an experimental package to attempt a one click isolated install and will download and use the pixi package manager, and because the repository publishes no GitHub releases. Verify three things before you rely on it: which of the three install routes you are comfortable with, since the readme orders them by speed and then reliability and calls the manual one the most reliable, whether the automatic install will work in your environment given that it fetches the pixi package manager, and how large the model download is, since the models are pulled from Hugging Face on first use. The licence is MIT, the package version in the manifest is 0.2.19, and the last push was on 2026-08-24.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. PozzettiAndrea/ComfyUI-TRELLIS2 on GitHub
4. README