Research map / 3D vision / Gaussian splatting

Hardware-Rasterized Ray-Based Gaussian Splatting

CVPR 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a hardware-rasterized renderer for ray-based 3D Gaussian Splatting (RayGS) that encloses each primitive's support with a 3D quad, plus an approximate MIP formulation for alias-free rendering across scales.

Problem

Rendering RayGS (ray-traced 3D Gaussian volume rendering) models is too slow for high frame-rate applications such as VR/MR on consumer hardware, since ray-based formulations lack the straightforward hardware rasterization mapping of standard 3DGS; RayGS also suffers from aliasing when rendering at

Approach

Instead of enclosing the projected support with a 2D screen-space quad (which can be a half-hyperbola for RayGS), the method isomorphically maps the boundary of the support to a unit circle and places a minimum-area quad in 3D space, then interpolates a quantity in the fragment shader to recover the RayGS divergence; it also derives a pixel-area-smoothed (MIP) opacity modulated by a pixel-dependen

Key results

• On MipNeRF360 and Tanks&Temples scenes, VKRayGS renders on average about 40x faster than the GOF CUDA renderer on an RTX2080 (e.g., bicycle 4 FPS vs 177 FPS).
• Quality is slightly lower than GOF (e.g., bicycle PSNR 25.47 vs 25.31; barn 28.99 vs 28.30), described as an implementation misalignment rather than a consequence of the method.
• The MIP formulation removes aliasing at distant views without the thickening artifacts of naive MIP-like smoothing, at negligible computational overhead.

Measured effects reported in the paper

Surprising findings

• VKGS (hardware-rasterized 3DGS) is on average about 2x faster than the INRIA CUDA 3DGS renderer, with quality metrics largely unchanged.

Limitations

• Assumes near-plane clipping is disabled; quads intersecting the near plane produce visible discontinuities.
• Quality is slightly below GOF on several scenes, attributed to pipeline misalignments.
• Only accelerates test-time rendering; training remains CUDA-based and not accelerated by hardware rasterization.

Experimental setup

Datasets: MipNeRF360 Tanks and Temples

Benchmarks: MipNeRF360 Tanks and Temples

Metrics: PSNR SSIM LPIPS FPS

Baselines: GOF (Gaussian Opacity Fields) CUDA renderer 3DGS (INRIA CUDA renderer) VKGS MSAA

Builds on: 3D Gaussian Splatting RayGS (Gaussian Opacity Fields) VKGS MIP-Splatting

Closest papers by meaning

• DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds (CVPR 2025 · similarity 0.92)
• FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.91)
• 6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering (ICLR 2025 · similarity 0.91)
• Anti-Aliased 2D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
• IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
• Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.91)