-
-
Save shawngraham/9899e77229ddae7ba8068aca083fe0f8 to your computer and use it in GitHub Desktop.
# https://github.com/nerfstudio-project/nerfstudio/issues/2438 | |
conda create --name nerfstudio -y python=3.8 | |
conda activate nerfstudio | |
python -m pip install --upgrade pip | |
pip uninstall torch torchvision functorch tinycudann | |
pip install torch==2.0.1 torchvision==0.15.2 | |
git clone https://github.com/nerfstudio-project/nerfstudio.git | |
cd nerfstudio | |
pip install --upgrade pip setuptools | |
pip install -e . |
bah.
ok, I got it running; it's eating nearly 13 gb of ram on 'poster', I dropped the 'mixed-precision' flag, and I don't have the viewer running. It's slow as hell, but... it's working?
ok, turning on the viewer kills everything. Alright. However - with colmap installed ($ brew install colmap
), I can feed my folder of images like so: ns-process-data images --data data/nerfstudio/test --output-dir data/nerfstudio/testoutput
and then use the version of nerfstudio running in the browser https://colab.research.google.com/github/nerfstudio-project/nerfstudio/blob/main/colab/demo.ipynb to nerfify things. Just zip the processed folder (which contains transforms.json), and then use the 'use polycam' button to upload things. Then make sure it unzips properly so that the contents of the unzipped folder are in data/nerfstudio/custom data.
(ie, nerfstudio runs on my M1 mini, but excruciatingly slowly.)
ns-process-data video --data data/nerfstudio/testfilm.mp4 --output-dir data/nerfstudio/videooutput
longer videos, with slow movement around the object, seem to get best results.
PYTORCH_ENABLE_MPS_FALLBACK=1 ns-train nerfacto --data data/nerfstudio/poster --machine.device-type mps --mixed-precision False