A high-performance, differentiable implementation of 3DGS — from video to trained Gaussians, fully in Rust.
Complete analytic backward pass through the entire rendering pipeline — no autograd dependency. Validated against finite differences.
Optional wgpu compute shaders for fast inference. CPU fallback with Rayon parallelism for training on any machine.
Tokio-based async orchestration. Non-blocking ffmpeg/COLMAP subprocesses, parallel rendering, async checkpoint saves.
Reads/writes standard 3DGS .ply files and compact .splat format. Compatible with existing viewers and tools.
End-to-end: video frame extraction → COLMAP SfM → Gaussian init → Adam training → .ply export.
Optional PyO3 bindings via maturin. Use the full pipeline from Python with an async API.
# Linux / macOS bash <(curl -fsSL https://raw.githubusercontent.com/mfaeezshabbir/rgsplat/main/scripts/install.sh) # Windows (PowerShell as Admin) powershell -ExecutionPolicy Bypass -File https://raw.githubusercontent.com/mfaeezshabbir/rgsplat/main/scripts/install.ps1
rgsplat --input my_scene.mp4 --output ./output
See the full documentation for all options.