mlx-splat: 3D Gaussian Splatting Renderer for Apple Silicon

A minimal, render-only implementation of 3D Gaussian Splatting using MLX + Metal. This is a personal experimental project for learning and exploration.

Note: This is a rendering-focused implementation. Training and optimization code has been removed.

Acknowledgments

This project is inspired by and builds upon excellent work in the 3D Gaussian Splatting community:

• 3D Gaussian Splatting - Original paper and CUDA implementation by Kerbl et al.
• gsplat - Fast CUDA kernels and efficient Python bindings
• OpenSplat - Open-source training pipeline

This implementation explores 3DGS rendering on Apple Silicon using MLX and Metal, while maintaining compatibility with standard PLY file formats from the above projects.

Installation

uv sync
source .venv/bin/activate  # if needed

Usage

CLI Renderer

uv run splat-render \
  --ply data/your_scene.ply \
  --cameras data/cameras.json \
  --camera_id 0 \
  --output outputs/render.png

Main options: --algorithm (rasterization), --tile_algorithm, --tile_size, --sigma_cut, --max_gaussians_per_tile, etc.

Interactive Viewer

uv run splat-viewer \
  --ply data/your_scene.ply \
  --cameras data/cameras.json

Orbit controls with mouse, keyboard navigation, and real-time parameter adjustment.

Python API (Minimal Example)

import mlx_splat

gaussians = mlx_splat.load_ply("data/your_scene.ply")
cameras = mlx_splat.load_cameras("data/cameras.json")
camera = cameras[0]

image = mlx_splat.render_gaussians(
    gaussians,
    camera,
    background=(0.0, 0.0, 0.0),
    raster_algorithm="metal",
    tile_algorithm="global",
    tile_size=(16, 16),
)

Project Structure

• src/mlx_splat/ Rendering core (core.py, io_utils.py, projection.py, tile_binning.py, rasterization*.py, renderer.py)
• examples/ Example scripts and interactive viewer
• tests/ Unit and integration tests

Testing

uv run pytest tests/ -v

Features

• ✅ MLX/Metal implementation - Runs on Apple Silicon GPUs
• ✅ Multiple rasterization backends - Metal and reference implementations
• ✅ Flexible tile binning - Global and local strategies for different scene sizes
• ✅ Standard PLY format - Compatible with models from 3DGS, gsplat, OpenSplat
• ✅ Interactive viewer - Real-time orbit controls with DearPyGUI
• ✅ Comprehensive tests - Unit, integration, and property-based testing

Limitations

• Rendering only - No training or optimization (use OpenSplat or original 3DGS for that)
• Apple Silicon only - Requires Metal-capable Mac (M1/M2/M3/M4)
• Experimental - API may change in future versions

Contributing

Contributions are welcome! Please feel free to submit issues or pull requests.

License

MIT License - see LICENSE file for details.