TRELLIS.2 Local Setup: Hardware, CUDA, and ComfyUI

Local setup is a systems project

Running TRELLIS.2 locally can remove a hosted per-job credit and give you control over data and integration. It also makes you responsible for the GPU, driver, CUDA toolkit, Python environment, model files, security review, and upgrades. Plan the work as a systems project, not a single install command.

The official upstream project is Microsoft TRELLIS.2. ComfyUI integrations are community projects with their own code, releases, and security posture. They are not automatically supported by Microsoft or by trellis-2.com.

What the official project expects

• Linux and an NVIDIA GPU with at least 24 GB of memory
• Verification on A100 and H100 hardware
• Python 3.8 or later and a Conda environment
• PyTorch 2.6.0 with CUDA 12.4
• flash-attn by default, or xformers as the documented fallback

Read the current README at the commit you plan to install. Requirements and commands can change, and older TRELLIS instructions do not transfer to TRELLIS.2.

Setup order that keeps failures isolated

1. Record the operating system, driver, CUDA toolkit, Python, PyTorch, model revision, and GPU model.
2. Clone the official repository recursively and follow its current setup instructions.
3. Download the official model from the linked model repository and record hashes.
4. Run the official example at the lowest supported resolution.
5. Only after that, evaluate a community ComfyUI wrapper.
6. Inspect every wrapper dependency and binary wheel before installing it.
7. Run a second time after a process restart to expose cache assumptions.

The full download guide covers provenance and the ComfyUI guide covers the wrapper boundary and troubleshooting.

Validation checklist

• The official pipeline succeeds on a small test image.
• Peak GPU and host memory are recorded.
• The GLB opens and the visible sides are inspected.
• Material channels and texture size are checked.
• A controlled failure releases temporary files and GPU memory.
• The wrapper and model revisions are pinned.
• Source-image and output licences are reviewed.

Do not call the setup production-ready after one successful sample.

When the hosted route is the better test

Use the hosted image-to-3D generator when you want one evaluation without installing CUDA or managing model files. It shows the selected resolution and credit cost before submission, tracks the job, and returns a GLB you can inspect in the free 3D viewer.

Use local setup when sustained volume, data control, or custom integration justifies the maintenance burden.

FAQ

Can I use ComfyUI with TRELLIS.2?

Yes, through a community wrapper. Treat it as separate software: review its code, dependencies, model revision, and licence before installing it.

Why does the setup fail in a CUDA step?

Check the exact driver, toolkit, CUDA_HOME, Python version, and package that failed. Mixing wheels from another CUDA or Python combination usually makes the problem worse.

How much VRAM do I need?

The official project currently documents at least 24 GB of NVIDIA memory. Memory use varies by resolution and loaded components; measure your own run.

Can I run this on macOS?

The official README's verified path is Linux with NVIDIA hardware. Do not assume a macOS workflow is supported without testing the current project and wrapper.

Read the official download steps

Open the TRELLIS.2 4B specification