A Gaussian Splatting viewer framework via GPU interprocess communication for real-time rendering in external applications (e.g., Unity, Blender).
 Unity plugin visualizing real-time 3D Gaussian Splatting results with interactive point clouds and 3D meshes. |
|
 Unity plugin visualizing real-time 4D Gaussian Splatting scenes with dynamic content. |
 Unity plugin visualizing real-time Gaussian avatar with 3D meshes. |
SplatBus is a viewer framework that enables zero-copy GPU memory sharing between Python-based Gaussian Splatting renderers and external viewer applications. By leveraging CUDA IPC (Inter-Process Communication) handles, SplatBus shares GPU memory directly across processes, eliminating expensive CPU-GPU data transfers and enabling real-time interactive visualization.
┌─────────────────────────────────────┐ ┌─────────────────────────────────────┐
│ Python Renderer │ │ Viewer (Unity/etc) │
│ (3DGS, 4DGS, GaussianAvatar...) │ │ │
│ │ │ │
│ ┌─────────────────────────────┐ │ │ ┌─────────────────────────┐ │
│ │ GaussianSplattingIPC │ │ │ │ Viewer Application │ │
│ │ Renderer │ │ │ │ │ │
│ └──────────┬──────────────────┘ │ │ └───────────┬─────────────┘ │
│ │ │ │ │ │
│ ┌──────────▼──────────────────┐ │ │ ┌───────────▼─────────────┐ │
│ │ SharedBuffer (Color) │◄───┼─────┼────┤ CUDA IPC Memory Map │ │
│ │ SharedBuffer (Depth) │ │ GPU │ │ (Zero-copy access) │ │
│ └──────────┬──────────────────┘ │ MEM │ └───────────▲─────────────┘ │
│ │ │ │ │ │
│ ┌──────────▼──────────────────┐ │ │ ┌───────────┴─────────────┐ │
│ │ IPCSocketServer (:6001) │────┼─────┼───►│ Receives IPC handles │ │
│ │ (sends CUDA handles) │ │ TCP │ │ & buffer metadata │ │
│ └─────────────────────────────┘ │ │ └─────────────────────────┘ │
│ │ │ │
│ ┌─────────────────────────────┐ │ │ ┌─────────────────────────┐ │
│ │ MessageSocketServer (:6000)│◄───┼─────┼────┤ Sends camera pose & │ │
│ │ (receives poses) │ │ TCP │ │ point cloud transforms │ │
│ └─────────────────────────────┘ │ │ └─────────────────────────┘ │
└─────────────────────────────────────┘ └─────────────────────────────────────┘
- Zero-Copy GPU Memory Sharing: Direct CUDA memory sharing via IPC handles without CPU involvement
- Dual Buffer Architecture: Separate color (RGBA32F) and depth (R32F) buffers for flexible rendering pipelines
- Bidirectional Communication: Real-time camera pose and point cloud transform updates between renderer and viewer
- CUDA Event Synchronization: Proper read/write synchronization across processes via CUDA events
- Easy Integration: Simple API designed for seamless integration with existing Gaussian Splatting codebases
- PyTorch >= 2.0.0 with CUDA support
- CUDA-capable GPU (the same GPU must be accessible by both renderer and viewer processes)
- Linux or Windows
- The Unity viewer uses D3D11 on Windows (instead of OpenGL on Linux). The native plugin handles this automatically via CUDA-D3D11 interop.
# From source
cd splatbus
pip install -e .
# Or directly
pip install .See instructions to run SplatBus with various Gaussian Splatting examples (3D/4D/Avatar).
- Unity viewer
- OpenGL viewer
- Blender viewer coming soon...
splatbus/
├── splatbus/ # Python package (renderer side)
│ ├── splatbus/ # Core library
│ │ ├── core/
│ │ ├── renderer.py # GaussianSplattingIPCRenderer
│ │ ├── client.py # GaussianSplattingIPCClient
│ │ └── camera.py # IPCCamera interface
│ └── examples/ # Python test scripts
├── clients/ # Viewer implementations
│ ├── Unity-client/ # Unity viewer plugin
│ ├── Blender-client/ # Blender viewer addon (coming soon)
│ └── OpenGL-client/ # OpenGL viewer
└── examples/ # Integration examples
├── 3dgs/ # 3D Gaussian Splatting
├── 4dgs/ # 4D Gaussian Splatting
└── gaussian-avatar/ # GaussianAvatar (mmlphuman)
# Terminal 1: Start the Python renderer (simulates Gaussian Splatting rendering)
python splatbus/examples/server_test.py
# Terminal 2: Start the Python simulated viewer (save the images on the disk)
python splatbus/examples/client_test.pyViewer options (Python test script):
python splatbus/examples/client_test.py \
--host 127.0.0.1 \
--ipc-port 6001 \
--msg-port 6000 \
--interval 0.1The examples/ directory contains standalone splatbus_renderer.py scripts for integrating SplatBus with various Gaussian Splatting implementations. See the examples README for detailed usage instructions.
3dgs/splatbus_renderer.py- Integration with 3D Gaussian Splatting4dgs/splatbus_renderer.py- Integration with 4D Gaussian Splattinggaussian-avatar/splatbus_renderer.py- Integration with mmlphuman
GaussianSplattingIPCRenderer Methods:
update_frame(color_data, depth_data, inverse_depth=True)- Update shared GPU buffers with new rendered frameupdate_view(view)- Update camera view from pose received from viewerupdate_gaussians(gaussians)- Update Gaussian positions from transform received from viewerclose()- Release resources and close connections
MIT License - see LICENSE for details.
If you find SplatBus useful for your research and applications, please cite using this BibTeX:
@misc{xu2026splatbusgaussiansplattingviewer,
title={SplatBus: A Gaussian Splatting Viewer Framework via GPU Interprocess Communication},
author={Yinghan Xu and Théo Morales and John Dingliana},
year={2026},
eprint={2601.15431},
archivePrefix={arXiv},
primaryClass={cs.GR},
url={https://arxiv.org/abs/2601.15431},
}