Python bindings for streamlined 3D Gaussian Splatting rendering.

• This is a more modularized and organized code of my previous c++ project vkgs.

Splatstream interactive viewer for "garden" scene.

Splatstream interactive viewer for MVP pre-trained model.

• Fast rendering: GPU-accelerated rasterization for high performance
  • Up to 2x faster than vk_gaussian_splatting viewer
  • Up to 1.3x faster than gsplat off-screen rendering
• Memory-efficient: resource reuse through double buffering
• Pure Vulkan implementation with minimal dependencies, no CUDA required
• Compatible with pre-trained models from the original Gaussian Splatting
• Interactive viewer

Install the stable version from PyPI:

$ pip install splatstream

Or install the development version from source:

$ pip install --no-build-isolation ./binding

import splatstream as ss

# Splats from .ply file
splats = ss.load_from_ply("my_splats.ply")

# Or, from numpy array
splats = ss.gaussian_splats(
    means,      # np.ndarray, (N, 3)
    quats,      # np.ndarray, (N, 4), wxyz convention.
    scales,     # np.ndarray, (N, 3)
    opacities,  # np.ndarray, (N)
    colors,     # np.ndarray, (N, K, 3), where K is one of [1, 4, 9, 16]
)

# Show in viewer
ss.show(splats)  # opens an interactive viewer

# Show in viewer with camera params
ss.show(
    splats,
    viewmats,          # np.ndarray, (..., 4, 4)
    Ks,                # np.ndarray, (..., 3, 3)
    width,             # int
    height,            # int
)

# Draw to numpy array
images = ss.draw(
    splats,
    viewmats,          # np.ndarray, (..., 4, 4)
    Ks,                # np.ndarray, (..., 3, 3)
    width,             # int
    height,            # int
    near=0.1, far=1e3  # Do not use too low near nor too high far, otherwise z-fighting
).numpy()              # np.ndarray, (..., H, W, 4), np.uint8, RGBA.

from PIL import Image
for i in range(len(images)):
  Image.fromarray(images[i, :, :, :3]).save(f"color_{i}.png")
  Image.fromarray(images[i, :, :, 3]).save(f"alpha_{i}.png")

• WASD/Space to move
• Q/E to roll
• Left drag to rotate
• Right drag to move
• L+R drag or wheel to zoom
• Ctrl wheel to dolly zoom (change FOV angle)

• GPU with Vulkan 1.4+ support (check compatibility)
• Up-to-date graphics drivers
• Python >= 3.10
• Supported platforms: Windows x64, Ubuntu x64, macOS arm64
• Monitor required for the viewer program

• No gradient computation; rendering only, not training
• Perfect pinhole cameras only
• Rendered images may differ slightly from gsplat (likely due to splat culling logic)
• No PyTorch integration

This project redistributes a modified subset of the Gaussian Splatting dataset.

The dataset is available here (Google Drive link).

See the LICENSE file in the dataset for details.

Download the datasets to the models/ directory.

To run the viewer with a pre-trained 3DGS scene:

$ python test/test_viewer.py --scene <scene>
$ python test/test_viewer.py --scene <scene> --no-camera

where <scene> is one of: bicycle, bonsai, counter, etc.

Tested on NVIDIA GeForce RTX 5080, Windows. FPS measure includes rendering time plus GPU-to-CPU transfer.

See bench for more details.

See DETAILS.md.

• VulkanSDK>=1.4.328.1 (set up environment variables per OS)
• cmake>=3.15 (install via pip install cmake or conda install conda-forge::cmake)
• python>=3.10

Feedback, feature requests, and issue reports are welcome!

• Device selection