Virtualized 3D Gaussians: Flexible Cluster-based Level-of-Detail System for Real-Time Rendering of Composed Scenes
(SIGGRAPH 2025 Conference Paper)
Xijie Yang, Linning Xu, Lihan Jiang, Dahua Lin, Bo Dai
All experiments were conducted on a Linux server running Ubuntu 22.04 with CUDA 11.8. Other versions may also be compatible.
conda create -y -n V3DG python=3.12
conda activate V3DG
pip install torch==2.6.0+cu118 torchvision==0.21.0+cu118 --index-url https://download.pytorch.org/whl/cu118
pip install plyfile==1.1 einops==0.8.1 tqdm==4.67.1 scipy==1.15.2 gsplat==1.4.0 opencv-python==4.11.0.86Place 3DGS assets under ./_3dgs-assets/. The donut asset is already included. Additional assets used in the paper are available on huggingface.
To build an asset (e.g., the donut asset at ./_3dgs-assets/donut/donut/point_cloud/iteration_30000/fused.ply:
# For quick testing: set simplification iteration to 16. Note: the first-time gsplat compilation takes a few minutes; build time ~10s.
python v3dg-build.py --ASSET_KIND donut --ASSET_NAME donut --ASSET_GSPLY_FILENAME point_cloud/iteration_30000/fused.ply --SIMPLIFICATION_ITERATION 16
# For default full build (640 iterations): build time ~100s.
python v3dg-build.py --ASSET_KIND donut --ASSET_NAME donut --ASSET_GSPLY_FILENAME point_cloud/iteration_30000/fused.plyThe build process generates a bundle of the asset in ./_v3dg-bundles/, which includes clusters at multiple levels of detail.
To see all available parameters for v3dg-build.py, refer to the VGBuildConfig class in ./libraries/cliconfigs.py.
To build all assets used in the paper, see the scripts in ./_scripts/ (./_scripts/v3dg-build.sh, ./_scripts/v3dg-build-ablations.sh, and ./_scripts/v3dg-build-not-rectified-scenes.sh).
Core code: ./libraries/classes.py
coarser_lod_layer = current_lod_layer.buildCoarserLodLayer() in Cluster.buildAllLodLayers() -> Clusters.buildCoarserLodLayer().
To render composed scenes with 3DGS instances, first prepare a layout description file. Examples are available in ./_layout-descriptions/, such as ./_layout-descriptions/_SingleDonut.json and ./_layout-descriptions/DonutSea.json.
Since bundles are saved with timestamps (e.g. donut_%y%m%d-%H%M%S_iter640_nginc4096_ncincg2), make sure bundle_names in the layout descriptions match accordingly.
To render scenes with vanilla 3DGS and our V3DG system:
# Test rendering a single donut at a fixed view.
python v3dg-select.py --LAYOUT_DESCRIPTION_JSON ./_layout-descriptions/_SingleDonut.json --TAU 2048 --SAVE_METRICS_AT_DIFFERENT_DISTANCES f
# Render the scene DonutSea (2048 donuts, 400 views). Full evaluation costs ~48 mins and reaches ~79 GB peak GPU memory (on NVIDIA A100, might not be the minimum need).
# Note: Scene composition is currently single-threaded and lacks a streaming module to optimize memory usage.
python v3dg-select.py --LAYOUT_DESCRIPTION_JSON ./_layout-descriptions/DonutSea.json --TAU 2048Renders (approx. 5 GB per scene) are saved in ./_output/, along with metrics recorded in _metrics.py and each filename.
To see all parameters for v3dg-select.py, refer to VGSelectConfig in ./libraries/cliconfigs.py.
To render all composed scenes used in the paper, check ./_scripts/ (./_scripts/v3dg-select.sh, ./_scripts/v3dg-select-ablations.sh, and ./_scripts/v3dg-select-comparisons.sh). For rendering parts in the supplementart video, check ./_layout-descriptions/_videos/.
Core code: ./v3dg-select.py vg_selected_gaussians = vg_layout_bundle.selectedGaussians(...) in VgSelect() -> VgBundle.selectGaussians().
The main branch contains the exact code used for the reproducibility of all experiments and results presented in the paper. The dev branch hosts ongoing development and improvements of this project.
@inproceedings{Yang2025V3DG,
author = {Yang, Xijie and Xu, Linning and Jiang, Lihan and Lin, Dahua and Dai, Bo},
title = {Virtualized 3D Gaussians: Flexible Cluster-based Level-of-Detail System for Real-Time Rendering of Composed Scenes},
booktitle = {ACM SIGGRAPH 2025 Conference Papers},
year = {2025},
doi = {10.1145/3721238.3730602},
url = {https://doi.org/10.1145/3721238.3730602},
publisher = {Association for Computing Machinery},
series = {SIGGRAPH '25}
}