Current base miner implementation is based on recently released TRELLIS 2 mesh generation model. TRELLIS.2 is a state-of-the-art large 3D generative model (4B parameters) designed for high-fidelity image-to-3D generation. It leverages a novel "field-free" sparse voxel structure termed O-Voxel to reconstruct and generate arbitrary 3D assets with complex topologies, sharp features, and full PBR materials.
To run this generator you will need a GPU with at least 48 GB of VRAM. It can work with GPUs from NVIDIA Blackwell family. You can run it on Geforce 5090 RTX if the generation settings are set to 512 resolution when you call run(...) method (see serve.py).
- latest docker package (we provide docker file in "docker" folder) or latest conda environment (we provide "conda_env.yml");
- NVIDIA GPU with cuda 12.8 support
- python 3.11
The code needs access to the gated model (commercially compliant) on huggingface: https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m
- Docker (building & pushing to remote register):
cd /docker
docker build --build-arg GITHUB_USER="" --build-arg GITHUB_TOKEN="" -t docker_name:docker-tag .
docker tag docker_name:docker-tag docker-register-path:docker-register-name
docker push docker-register-path:docker-register-name - Conda Env. (shell script will install everything you need to run the project):
# to install conda env
bash setup_env.sh
# to uninstall conda env
bash cleanup_env.sh- Docker (run locally):
docker run --gpus all -it docker_name:docker-tag bash
# outside docker
curl -X POST "http://0.0.0.0:10006/generate" -F prompt_image_file=@sample_image.png -o sample_model.glb
- Conda Env.:
# start pm2 process
pm2 start generation.config.js
# view logs
pm2 logs
# send prompt image
curl -X POST "http://0.0.0.0:10006/generate" -F prompt_image_file=@sample_image.png -o sample_model.glb
This model and code are released under the MIT License.