installation.md

Installation

A. Clone repository

# clone repository
git clone --recursive https://github.com/M3phist0/MetroGS.git
cd MetroGS

B. Create virtual environment

# create virtual environment
conda create -n metrogs python=3.10
conda activate metrogs

C. Install PyTorch

• We reimplemented the method on RTX 5090 due to an equipment upgrade in our group.

• Tested on PyTorch==2.10

• You must install the one match to the version of your nvcc (nvcc --version)

• For CUDA 12.8

  pip install torch torchvision -i https://pypi.tuna.tsinghua.edu.cn/simple -f https://mirrors.aliyun.com/pytorch-wheels/cu128
  

D. Install submodules

Use --no-build-isolation to resolve the error “ModuleNotFoundError: No module named 'torch'”.

# basic
pip install submodules/dist-2dgs --no-build-isolation
pip install submodules/simple-knn --no-build-isolation
# patchmatch | Note: submodules/propagation/setup.py "compute_xx" and "-D_GLIBCXX_USE_CXX11_ABI" must match to your_env_setting
pip install submodules/propagation --no-build-isolation
# apperance
pip install submodules/nvdiffrast --no-build-isolation
pip install submodules/tiny-cuda-nn/bindings/torch --no-build-isolation

Note: The -gencode=arch=compute_xx,code=sm_xx setting in submodules/propagation/setup.py must be adjusted to match your hardware environment.

diff-gaussian-rasterization is required by the base classes, even though we do not use it directly.

pip install submodules/diff-gaussian-rasterization --no-build-isolation

E. Install requirements

pip install -r requirements.txt

F. Download Prior Models

MoGe-2

NOTE: do not run pip install -r utils/MoGe/requirements.txt

# Mode: moge-2-vitl-normal
Download link: https://huggingface.co/Ruicheng/moge-2-vitl-normal/blob/main/model.pt
# Please download `model.pt` and move it to `utils/MoGe/checkpoints`

Pi3-Align

# Please follow the official script to download the weights:
cd pointmap/Pi3-Align 
bash scripts/download_weights.sh