Project Page | Paper | CUDA Rasterizer | Custom Dataset (15.7GB)

This is the official repository for "Nexels: Neurally-Textured Surfels for Real-Time Novel View Synthesis with Sparse Geometries".

First, clone the repository. Note that we have a number of submodules, so you should include the recursive flag.

git clone [email protected]:victor-rong/nexels.git --recursive

To install, create a conda environment.

conda create -n nexels python=3.8 -y
conda activate nexels

A suitable version of CUDA toolkit and PyTorch is needed. This codebase has been tested with CUDA 11.8.

conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit -y
pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

Our method uses the Instant-NGP architecture for the neural texture, which can be installed with

pip install --no-build-isolation git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

The other requirements, including the CUDA nexel rasterizer, can be installed via pip from requirements.txt.

pip install --no-build-isolation -r requirements.txt

Our method works on scenes preprocessed with COLMAP (see these instructions on setting up custom datasets). To train a nexels model with the recommended settings, run

python train.py -s $DATA_DIR -m $OUTPUT_DIR

To set the maximum number of primitives, you can set the --cap_max_init and --cap_max_final flags. The former determines the initial number of primitives and the latter is the maximum number of primitives by the end of training. Setting the initial amount to be half the final amount works consistently well. It may also be helpful to adjust --log_hash_table_size, which determines the amount of parameters used for the neural field. Increasing this generally improves results if memory is available.

The model can then be rendered across the train and test sets, as well as an estimated ellipse trajectory with

python render.py -s $DATA_DIR -m $OUTPUT_DIR

A browser-based viewer will be released soon.

Training scripts have been prepared to reproduce the paper's results on the datasets. Run

python scripts/eval_m360.py --start 0 --end 9 --data_dir $DATA_DIR --cap_max $CAP_MAX
python scripts/eval_custom.py --start 0 --end 4 --data_dir $DATA_DIR --cap_max $CAP_MAX
python scripts/eval_tntdb.py --start 0 --end 2 --data_dir $DATA_DIR --cap_max $CAP_MAX

to train nexel models on the MipNeRF-360, custom, and Tanks & Temples datasets with a maximum amount of primitives, $CAP_MAX.

This method was inspired by several great works. The codebase is built off of the original 3DGS code from Kerbl et al. We also used portions of gsplat from the Nerfstudio team and 2DGS by Huang et al.

If you find this repository useful in your projects or papers, please consider citing our paper:

@article{rong2025nexels,
    title={Nexels: Neurally-textured surfels for real-time novel view synthesis with sparse geometries},
    author={Rong, Victor and Held, Jan and Chu, Victor and Rebain, Daniel and
        Van Droogenbroeck, Marc and Kutulakos, Kiriakos N and Tagliasacchi, Andrea and Lindell, David B},
    journal={arXiv preprint arXiv:2512.13796},
    year={2025}
}