This repository contains the minimal code needed to run the 2DNS (2D neural splatting, proposed in Section 5.4 of Learning View-Dependent Splatting Kernels.) pretraining and benchmark pipeline.
mlp_pretrain.py: pretrains the neural renderer checkpoint.benchmark.py: runs dataset-level experiments and callstrain.py.train.py: main training entry point.arguments/,gaussian_renderer/,scene/,utils/,lpipsPyTorch/: runtime Python modules.submodules/beta_splat,submodules/diff-gaussian-rasterization,submodules/diff-neural-rasterization,submodules/diff-surfel-rasterization: vendored CUDA extension source code.submodules/fused-ssim,submodules/simple-knn, and eachglmdependency: Git submodules.
The tested setup uses Python 3.9, PyTorch 2.5.1 (cu124 wheel), a CUDA 12.x
compiler, and local CUDA extensions under submodules/.
Clone with submodules enabled:
git clone <repo-url> --recursive
cd 2dnsIf the repository was cloned without submodules:
git submodule update --init --recursiveconda env create -f environment.yml
conda activate 2dns
pip install -r requirements.txtUpdate the dataset paths in script/run_b.sh and script/mlp_pretrain.sh to match
your local setup:
source_root="/path/to/nerf_synthetic"
m360=/path/to/mipnerf360
tat=/path/to/tanks_and_temples
db=/path/to/deep_blending
ns=/path/to/nerf_syntheticExample:
python mlp_pretrain.py \
-s /path/to/nerf_synthetic/lego \
-m ./output/pretrain/lego_test \
--iteration 15000 \
--num_init_pts 1 \
--checkpoint_iterations 7000 15000 \
--test_iterations 100 1000 7000 15000 \
--decoder_size 8 \
--latent3d_size 10 \
--latent2d_size 10 \
--latentMat_size 10 \
--encoding_level 2 \
--num_hidden_layers 3 \
--sh_degree 3 \
--evalThe expected checkpoint is:
<pretrain_output>/neural_render/iteration_15000/neural_render.pth
python benchmark.py \
--output_path ./output/lego_test \
-ns /path/to/nerf_synthetic \
--scenes lego \
-p ./output/pretrain/lego_test/neural_render/iteration_15000/neural_render.pth \
--iterations 30000 \
--checkpoint_iterations 7000 15000 20000 30000 \
--test_iterations 7000 15000 20000 25000 28000 30000 \
--latent3d_size 10 \
--latent2d_size 10 \
--latentMat_size 10 \
--decoder_size 8 \
--num_hidden_layers 3 \
--sh_degree 3 \
--encoding_level 2 \
--neural_render_lr_init 0.00016 \
--neural_render_lr_final 0.0000016 \
--neural_render_lr_start_steps 2000 \
--neural_render_lr_max_steps 30000 \
--color_lr_init 0.001 \
--color_lr_final 0.000001 \
--color_lr_delay_steps 3000 \
--color_lr_delay_mult 0.1 \
--color_lr_start_steps 2000 \
--proj_hidden_size 64 \
--proj_hidden_num 3 \
--color_hidden_size 128 \
--color_hidden_num 4 \
--opacity_lr 0.05 \
--densify_until_iter 25000 \
--color_mode sbTo reproduce the full paper benchmark across all datasets, specify paths to each
dataset directory. Omit --scenes to run all scenes automatically.
python benchmark.py \
--output_path ./output/full_benchmark \
-ns /path/to/nerf_synthetic \
-m360 /path/to/mipnerf360 \
-tat /path/to/tanks_and_temples \
-db /path/to/deep_blending \
-p /path/to/pretrained_checkpoint.pth \
--iterations 30000 \
--checkpoint_iterations 7000 15000 20000 30000 \
--test_iterations 7000 15000 20000 25000 28000 30000 \
--latent3d_size 10 \
--latent2d_size 10 \
--latentMat_size 10 \
--decoder_size 8 \
--num_hidden_layers 3 \
--sh_degree 3 \
--encoding_level 2 \
--neural_render_lr_init 0.00016 \
--neural_render_lr_final 0.0000016 \
--neural_render_lr_start_steps 2000 \
--neural_render_lr_max_steps 30000 \
--color_lr_init 0.001 \
--color_lr_final 0.000001 \
--color_lr_delay_steps 3000 \
--color_lr_delay_mult 0.1 \
--color_lr_start_steps 2000 \
--proj_hidden_size 64 \
--proj_hidden_num 3 \
--color_hidden_size 128 \
--color_hidden_num 4 \
--opacity_lr 0.05 \
--densify_until_iter 25000 \
--color_mode sbThe four supported datasets are:
- NeRF Synthetic (
-ns): chair, drums, ficus, hotdog, lego, materials, mic, ship - Mip-NeRF360 Outdoor (
-m360): bicycle, flowers, garden, stump, treehill - Mip-NeRF360 Indoor (
-m360): room, counter, kitchen, bonsai - Tanks & Temples (
-tat): truck, train - Deep Blending (
-db): drjohnson, playroom
Pass only the datasets you have available. Results are collected in a summary table
under --output_path.
Cite as below if you find this repository is helpful to your project:
@inproceedings{ding2026kernel,
title = {Learning View-Dependent Splatting Kernels},
author = {Huakeng Ding and Zhangpeng Liu and Fan Pei and Kun Zhou and Hongzhi Wu},
booktitle = {SIGGRAPH 2026 Conference Papers},
year = {2026}
}
We have intensively borrowed code from 2D Gaussian Splatting, 3D Gaussian Splatting, MCMC Gaussian Splatting, and Beta Splatting (beta color model). Many thanks to the authors for sharing their codes.
