Show HN: 3DGS implementation in Nvidia Warp: clean, minimal, runs on CPU and GPU
5 days ago
2
3DGS implementation in NVIDIA Warp — clean, minimal, and runs on CPU or GPU.
✅ CPU & GPU with Zero Hassle
Built with NVIDIA Warp, the same code runs seamlessly on both CPU and GPU — no need to deal with CUDA setup, driver issues, or device-specific kernels. Just flip one config line.
🧠 Learn Modern Graphics the Easy Way
Explore core concepts in differentiable rendering and parallel graphics programming — no need for expensive GPUs or thousands of lines of boilerplate.
📦 Minimalist & Educational
This isn’t another massive codebase. It’s a clean, hackable implementation built for clarity — perfect for study, prototyping, or teaching yourself how Gaussian Splatting works.
# Download the Lego dataset
bash download_example_data.sh
# Render 3 Gaussian points – a minimalist example
python render.py
You should see 3 Gaussian points like:
# Train on Lego dataset (CPU by default)# For GPU training, change DEVICE in config.py to "cuda"
python train.py
├── train.py # Main training loop
├── render.py # Rendering script to validate outputs; confirms forward pass correctness
├── config.py # Configuration and training parameters
├── forward.py # 3DGS: Forward pass (reimplementation of graphdeco-inria/gaussian-splatting)
├── backward.py # 3DGS: Backward pass (reimplementation of graphdeco-inria/gaussian-splatting)
├── loss.py # Loss functions for training (includes depth loss, though unused in this repo)
├── scheduler.py # Learning rate scheduler
├── utils/
│ ├── camera_utils.py # Load camera intrinsics and extrinsics from training data
│ ├── point_cloud_utils.py # Point cloud I/O utilities (e.g., saving to .ply)
│ ├── math_utils.py # General math utilities (e.g., transformation matrices)
│ └── wp_utils.py # Warp utilities for math operations and device transfer
└── data/ # Contains the NeRF-synthetic 'Lego' dataset
forward.py and backward.py are based on graphdeco-inria/gaussian-splatting. The original pure CUDA version is now reimplemented in Nvidia Warp, easy to understand, set up, and run.
Densification and pruning logic is based on yzslab/gaussian-splatting-lightning, but restructured here with minimal data preparation and simplified training logic.
Improve performance — This implementation focuses on correctness and clarity over speed; there's room for Warp kernel optimization.
Filter inactive points — The saved .ply files include many points that do not contribute to rendering. A better pruning strategy or visibility check is needed.
.png)
5 days ago
2
