TRELLIS.2-ROCm

Info:

This is a fork of TRELLIS.2 that enables running 3D model generation on AMD GPUs using ROCm.
The script uses HIPIFY, and most libraries are patched on the fly; therefore, make sure you have the same versions of ROCm and HIPIFY installed.
If you want a reliable test environment, it is recommended to use a Podman container (https://github.com/Mateusz-Dera/ROCm-AI-Installer ).

Note

An automatic installer for the script above will also be added.

The script uses nvdiffrast-hip from https://github.com/CalebisGross/TRELLIS-AMD

The script applies patches dynamically:

• https://github.com/JeffreyXiang/nvdiffrec.git
• https://github.com/JeffreyXiang/CuMesh.git
• https://github.com/JeffreyXiang/FlexGEMM.git

Replaced:

• RMBG-2.0 by https://github.com/plemeri/InSPyReNet (transparent-background)

Original repository: https://github.com/microsoft/TRELLIS.2
Original README: https://github.com/Mateusz-Dera/TRELLIS.2-ROCm/blob/main/ORGINAL_README.md

Note

Core model generation is functional, but I'm in the process of replacing certain modules. Consequently, features such as the model preview or proper texturing are currently unavailable.

Test platform:

Name	Info
CPU	AMD Ryzen 9 9950X3D
GPU	AMD Radeon 7900XTX
RAM	64GB DDR5 6600MHz
Motherboard	Gigabyte X870 AORUS ELITE WIFI7 (BIOS F8)
OS	Debian 13.2
Kernel	6.12.57+deb13-amd64

Instalation:

1. Install uv:

sudo apt -y install pipx
pipx install uv

2. Install libjpeg-dev

sudo apt install -y libjpeg-dev

3. Set the GPU architecture (If you are not using Podman):

export ROCM_PATH="/opt/rocm"
export HSA_OVERRIDE_GFX_VERSION="11.0.0" # RADEON 7900 XTX
export TARGET_GFX="gfx1100" # RADEON 7900 XTX
export PYTORCH_ROCM_ARCH="gfx1100" # RADEON 7900 XTX

3. Run setup.sh:

bash ./setup.sh

After installation:

1. Activate venv
2. Go to https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m
3. Click “Agree and access repository” to accept the license.
4. Activate your Hugging Face token in the terminal.

Run app:

source .venv/bin/activate
python ./app.py