Before you start

Prepare a clean environment, compile the required CUDA components, download the official checkpoint, and validate the smallest example before changing settings.

Prerequisites

  • Linux
  • NVIDIA GPU with at least 24 GB VRAM for the official path
  • CUDA toolkit
  • Conda
  • Git and Git LFS familiarity

Step by step

Confirm the upstream baseline

Read the current official prerequisites and match the recommended Python, PyTorch, and CUDA versions before creating the environment.

The GPU driver and CUDA toolchain are visible from the shell.

Clone with submodules

Use the official repository and initialize its submodules. Avoid mixing dependency instructions from first-generation TRELLIS.

Build the minimal environment

Run the repository setup for the basic packages and only the acceleration backends required by your GPU.

Download the official checkpoint

Use the exact TRELLIS.2-4B model identifier and retain the model card and license with the environment record.

Run the upstream example

Start from the supplied example image and default settings. Confirm both preview frames and GLB export before testing custom inputs.

The GLB opens and reports geometry plus material data.

Known trade-offs

  • Do not present community Windows or low-VRAM instructions as official support
  • CUDA extension builds are sensitive to compiler and toolkit mismatches
  • Keep generated assets separate from model and code license records

Sources

Official TRELLIS.2 installation ↗