This repository provides a pure PyTorch implementation of 3D Gaussian Splatting, targeting researchers and developers interested in efficient and high-quality 3D scene reconstruction and rendering. It offers a direct PyTorch alternative to existing C++/CUDA-based implementations, enabling easier integration and modification within the PyTorch ecosystem.

How It Works

The implementation leverages PyTorch for all computations, including the core splatting operations. It utilizes a tile-based rendering approach, processing scenes in 64x64 tiles to optimize performance within a Python loop, a departure from the 16x16 tiles used in the original Gaussian Splatting implementation. This design choice aims to balance computational efficiency with ease of use and PyTorch integration.

Quick Start & Requirements

• Install: conda env create -f environment.yml and conda activate torch-splatting
• Requirements: Python, PyTorch, CUDA (implied by RTX 2080Ti mention).
• Setup: Training time is approximately 2 hours for 512x512 resolution at 30k iterations on an RTX 2080Ti.

Highlighted Details

• Pure PyTorch implementation for easier integration.
• Tile-based rendering with 64x64 tiles for Python loop optimization.
• Achieves ~39 PSNR on synthetic data with 16384 Gaussians.

Maintenance & Community

No specific community channels, roadmap, or notable contributors are mentioned in the README.

Licensing & Compatibility

The README does not specify a license.

Limitations & Caveats

The number of 3D Gaussians is fixed at 16384. The implementation is presented as a work in progress ("Stay Tuned").