Great news for those following advancements in 3D Gaussian Splatting— the code for Efficient Perspective-Correct 3D Gaussian Splatting Using Hybrid Transparency (HTGS) is now publicly available! We covered the paper back when it was published, last year.
HTGS builds on previous 3DGS research by tackling popping artifacts and significantly boosting performance, with reported rendering speed increases of up to 2x, depending on the scene. This work introduces a perspective-correct projection method that eliminates depth-sorting issues and improves rendering stability without costly matrix inversions. The authors also implement a hybrid transparency model, ensuring that depth-blending is both visually accurate and computationally efficient.
The method refines key areas of 3DGS by leveraging Plücker coordinates and a novel core/tail blending approach to balance quality and real-time performance.
Key Features of HTGS
Perspective-Correct Projection: Ensures stable and accurate splat evaluation, particularly in challenging depth variations.
Hybrid Transparency Blending: Splits blending into core (top K fragments) and tail contributions, improving real-time rendering performance.
Optimized CUDA Kernels: Designed for speed and efficiency, reducing computational overhead.
Multiple Blending Modes:
HYBRID_BLEND: Blends top K fragments using alpha blending; remaining fragments use order-independent transparency.
ALPHA_BLEND_FIRST_K: Only the K foremost fragments are blended, discarding the rest.
ALPHA_BLEND_GLOBAL_ORDERING: Sorts Gaussians based on their Z-depth for approximate blending.
OIT_BLEND: Fully order-independent transparency blending for all fragments.
How to Get Started
The code is provided as an extension to the NeRFICG framework, which is required for setup. The repository includes:
Training and inference scripts compatible with NeRFICG.
Optimized CUDA implementations for efficient rendering.
Example configuration files for datasets like Mip-NeRF360 and Tanks & Temples.
Export support for .ply files (note: correct rendering requires ray-based evaluation of 3D Gaussians).
The authors have tested the implementation on an Nvidia RTX 4090 with Ubuntu 22.04 and optimized CUDA Toolkit versions. Be mindful that CUDA Driver Version 560 shows a 20% performance regression, so earlier versions are recommended for optimal results.
With the release of the HTGS code, developers and researchers can now experiment with these enhancements firsthand. As Radiance Fields continue to evolve, approaches like HTGS show how much room remains for optimization and innovation.
Check out the official code repository and start testing HTGS for yourself!
