Last night a flurry of updates got pushed into nerfstudio. The most exciting of which is a pytorch implementation of Zip-NeRF.

Zip-NeRF is a strong NeRF method from the Google Research team. Once installed, it takes roughly 55 minutes to train on my computer.

Zip-NeRF is the latest 3rd party method to receive support within the Berkeley based studio and has been hotly request, since its paper's release. Late last year, Google researcher and paper author Jon Barron released the code to both CamP and Zip-NeRF. This implementation is different from the official release is this one is built on PyTorch.

In order to install Zip-NeRF, please follow the below instructions.

First, install nerfstudio and its dependencies. Then run:

pip install git+https://github.com/SuLvXiangXin/zipnerf-pytorch#subdirectory=extensions/cuda
pip install git+https://github.com/SuLvXiangXin/zipnerf-pytorch

Finally, install torch_scatter corresponding to your Cuda version(https://pytorch-geometric.com/whl/torch-2.0.1%2Bcu118.html).

In addition to this change, other commits featuring the following were also added.

• Add transparency carving to splatfacto (#2889)

Before and After

• Switch default matching method for videos to 'sequential' in COLMAP (#2912)

• Fix an issue where gradient accumulation could not be passed as argument due to a type error. (#2910)

• Implement new rasterization mode using opacity compensation factor

gsplat also received an update to V0.1.5, with the following additions:

• Improved pytorch implementation of project_gaussians inside _torch_impl.py

• Screen space blur compensation for rasterizing tiny gaussians from Mip-Splatting

• Bugfixes