MiniGSRenderer

A minimal Gaussian Splat renderer with both CPU and CUDA paths, plus interactive viewers.

Quick Start

# Configure and build (Windows example)
cmake -S . -B build
cmake --build build --config Release

# Binaries are in build/Release

Apps

• MiniGSRenderer.exe: Offline renderer (CPU or CUDA) saving PNG output.

  • Usage:

    MiniGSRenderer.exe --scene <file.ply> --output <out.png> [--camera_config cameras.json] [--cam_id N] [--cuda]

  • Examples:

    MiniGSRenderer.exe --scene train.ply --output out.png --camera_config cameras.json --cam_id 2
    MiniGSRenderer.exe --scene train.ply --output out_cuda.png --camera_config cameras.json --cam_id 2 --cuda

• viewer_main.exe: Interactive viewer with real-time reprojection and CUDA PBO display.

  • Usage:

    viewer_main.exe <file.ply> [camera_config.json] [camera_id]

  • Controls: WASD move, Q/E up/down, Shift for speed, Mouse to look, ESC to quit.

• viewer_progressive.exe: Interactive progressive viewer that reveals splats from the screen center outward.

  • Usage (same args as viewer_main):

    viewer_progressive.exe <file.ply> [camera_config.json] [camera_id]

Camera Config (JSON)

Cameras are stored as a top-level array. Each entry has id, width, height, position (3 floats), rotation (3×3 rows, camera-to-world, row-major), and intrinsics fx, fy.

[
  {
    "id": 2,
    "img_name": "00017",
    "width": 1959,
    "height": 1090,
    "position": [-0.77375337, -0.33642719, -2.93589694],
    "rotation": [
      [0.99988134, 0.01374238, -0.00696055],
      [-0.01426837, 0.99651294, -0.08220929],
      [0.00580653, 0.08229885, 0.99659078]
    ],
    "fy": 1164.6601,
    "fx": 1159.5881
  }
]

Notes:

• rotation is camera-to-world in row-major format; the viewer constructs c2w and sets view = inverse(c2w).
• proj is built from fy, width, and height.
• The interactive viewers also sync the gs::FpsCamera position, yaw/pitch (derived from forward), and FOV from fy.

CPU vs GPU Rendering

• CPU Path

  • Projects 3D Gaussian splats to screen splats, sorts for visibility, and rasterizes on the CPU.
  • Deterministic and simple to debug; suitable for offline rendering.

• CUDA Path

  • Uploads scene-static data (positions, SH colors, etc.) to the GPU.
  • In MiniGSRenderer.exe, renders to a float RGB buffer then saves PNG.
  • In viewer_main.exe, renders directly into a CUDA-mapped PBO and updates an OpenGL texture for display.
  • Offers significantly higher throughput for large splat sets.

Build Details

• Requires: CMake, a C++17 compiler, GLM, GLFW, GLAD, CUDA toolkit (for GPU path), and nlohmann_json.
• Windows builds place executables in build/Release.

License

MIT License. See LICENSE.