Three ComfyUI TRELLIS node packs compared: install routes, node coverage and who each one suits

Microsoft Research released TRELLIS.2 in December 2025 under an MIT licence, and ComfyUI wrappers appeared within days. Three packs get most of the attention: visualbruno/ComfyUI-Trellis2 and PozzettiAndrea/ComfyUI-TRELLIS2 wrap TRELLIS.2, while if-ai/ComfyUI-IF_Trellis wraps the earlier TRELLIS 1.x line. All three land in the same custom_nodes folder and all three write a mesh file, but they ask for very different amounts of work before the first render.

This comparison looks at the four things that decide whether a pack survives contact with your machine: how the install is delivered, which compiled CUDA extensions it needs, what the node set covers once a mesh exists, and how recently the repository moved. Star counts appear as rough bands only, because they change every week, and every figure below is worth re-checking on the repository page before you spend an afternoon on it.

What a TRELLIS node pack actually wraps

None of these projects train a model. TRELLIS.2 is a 4-billion-parameter network that predicts a compact structured latent, called O-Voxel by Microsoft, and decodes it into geometry plus base colour, roughness and metallic maps from a single image. A ComfyUI pack loads those checkpoints, wires them into the graph and decides how much of the surrounding pipeline arrives as nodes: preprocessing and background removal, sparse-structure and shape sampling, texturing, remeshing and hole filling, then export. if-ai wraps the earlier TRELLIS 1.x line, a different model that also decodes Gaussian and radiance-field representations, so comparisons only hold inside one model version.

The three packs at a glance

The dependency surface, where installs actually fail

visualbruno/ComfyUI-Trellis2: bundled wheels and the largest node set

This pack treats installation as a supply problem. Prebuilt Windows wheels cover the parts that normally need a compiler, namely CuMesh, nvdiffrast, nvdiffrec_render, FlexGEMM and o_voxel, sorted into per-Torch folders with cp311 builds for Torch 2.7 and 2.8 and matching python_embeded commands for the portable version. Paste the five pip lines, install requirements.txt, then handle one more thing: TRELLIS.2 conditions on DINOv3, so the gated facebook/dinov3-vitl16-pretrain-lvd1689m has to sit in ComfyUI/models/facebook/dinov3-vitl16-pretrain-lvd1689m. The author lists Windows 11, Python 3.11 and Torch 2.7.0 with cu128 as the tested setup. The node set reaches past generation into cleanup: hole filling through Meshlib and CuMesh, quad remeshing, vertex welding, trimesh simplification, batch export, multiview generators, multi-view texturing and FP8 checkpoints.

Where the bundled wheels stop fitting

The changelog is the honest part of that repository: it added Python 3.13, Torch 2.10 and CUDA 13.1 wheels in February 2026, which also tells you the matrix is finite. If your combination falls outside it, you build CuMesh, o_voxel and FlexGEMM from the author's own forks, and a community branch exists for one specific Blackwell failure. Outside the matrix this is the hardest of the three to install; inside it, the work is copying and pasting.

PozzettiAndrea/ComfyUI-TRELLIS2: isolated dependencies and fewer manual steps

Try this one first if you would rather not manage a Python environment yourself. The README orders three routes by reliability: search for TRELLIS2 in ComfyUI Manager and install the highest version listed, add the Git URL through Manager, or clone it and run pip install -r requirements.txt followed by python install.py. The one-click route uses comfy-env, an experimental helper that builds an isolated environment with the pixi package manager. Isolation is the point here: the compiled geometry libraries live outside your ComfyUI environment, so the next node pack you install is less likely to break a working graph. The costs are a second package manager on disk and a slower first install.

if-ai/ComfyUI-IF_Trellis: the TRELLIS 1.x source build

The oldest of the three wraps TRELLIS 1.x, published in December 2024, and it is the only one that decodes radiance fields and 3D Gaussians alongside meshes. That still matters for view synthesis and for pipelines that want splats rather than triangles. Its README asks for an NVIDIA GPU with at least 8 GB, says the upstream code was tested only on Linux, and states plainly that the pack is not compatible with ComfyUI's portable embedded Python. Its last commits are from March 2025, so it predates TRELLIS.2 and everything built on top of it.

1. Install the CUDA toolkit the upstream guide names inside a conda or micromamba environment: 11.8 and 12.2 are the tested versions, and the repository pins 12.4.
2. Compile and install the three geometry extensions: diffoctreerast, the diff-gaussian-rasterization submodule from mip-splatting, and nvdiffrast.
3. Install kaolin from NVIDIA's index for Torch 2.4.0 with CUDA 12.1, and utils3d from the exact commit the README names.
4. Clone with --recurse-submodules, because the node pack fails at import without the submodules in place.
5. Start ComfyUI from that environment rather than the portable build; the open issues show missing Preview3D and Load3D node types when the environment is wrong.

Star counts and commit dates are weak signals

The licence surface belongs in the install decision

One dependency deserves a look before commercial delivery. The reference TRELLIS.2 pipeline rasterises and renders through NVIDIA's nvdiffrast and nvdiffrec, both released under a source-code licence that limits use to non-commercial research and evaluation, while the model's own code and weights are MIT. A pack that ships or compiles those libraries carries the question with it, and the prebuilt wheel list in one of these repositories includes nvdiffrast by name. ComfyUI's core 3D nodes take a different route and implement post-processing in plain PyTorch and SciPy, so if licence clarity matters more than node count, check what your build actually ships and read the licences yourself.

Choosing by situation rather than by ranking

• Windows plus portable ComfyUI, and your Torch version appears in the wheel folders: visualbruno's pack, after verifying the version match.
• You want ComfyUI Manager to resolve dependencies and you accept an isolated environment: PozzettiAndrea's pack.
• You already run ComfyUI from conda or micromamba on Linux and compile CUDA extensions without flinching: if-ai's pack, pinned to the commits it names.
• You need Gaussian or radiance-field output rather than a mesh: only if-ai's pack covers it, and it has been quiet for about a year.
• You need a clean licence trail for client work: settle the renderer question before you pick, because that decides the answer.

If the deliverable is an asset rather than a node stack, a hosted route removes most of this argument. Open3DAI runs one local procedural generation engine and can connect optional third-party engine adapters with your own API keys, so the same prompt or image can go through several engines and be compared side by side before export to GLB, glTF, OBJ, STL, PLY or 3MF; text-to-3D, image-to-3D, multiview input and AI texturing are covered, alongside 48 format converters, 12 browser tools and a REST API. It does not do photogrammetry, rigging, animation, auto-retopology or UV unwrapping, it does not produce CAD-grade parametric solids, and it does not promise a watertight mesh straight out of the export, so Blender or Meshmixer still has a place in the chain.

Frequently Asked Questions

Run it yourself in Open3DAI

Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.