Stochastic rendering method reduces noise in 3d gaussian splatting
SteadySplats: Resampling of Low-Variance Gaussians for High-Fidelity Stochastic Rendering
Computer Vision and Pattern RecognitionGraphics
Summary
Rendering images from 3D Gaussian Splatting models often results in noisy pictures that take a long time to look clean. The authors developed a way to reduce this noise by improving both the model and the rendering process. They use smart techniques to reuse information across space and time during rendering and introduce a color smoothing method during training. This makes the images clearer much faster and keeps them stable when the camera moves.
What this means in practice
- •For graphics software developers: Create rendering engines that produce less noisy images from 3D Gaussian Splatting models faster and with smoother camera motion.
- •For visual effects studios: Generate high-quality stochastic transparent effects with reduced noise and faster convergence, improving production efficiency.
Authors
Felix Windisch, Thomas Köhler, Lukas Radl, Chris Wyman, Georgios Kopanas, Bernhard Kerb, Markus Steinberger
Abstract
Stochastic order-independent transparency enables efficient and elegant rendering of primitive-based radiance fields like 3D Gaussian Splatting models, but remains impractical due to the inherent visible noise in the output. We propose a principled approach to minimize high-frequency noise, addressing its sources at the representation and image synthesis level. During stochastic rendering, our history-based spatial resampling scheme drastically accelerates image convergence, while temporal importance resampling ensures coherence under camera movement. During training, a color regularizer implicitly reduces the variance along view rays in the 3DGS models. With these properties, our optimized, Vulkan-based renderer effectively mitigates output noise at low and high sample counts, achieving a substantial 13~dB PSNR increase in quality over previous stochastic methods at 1 sample per pixel and quickly converging to sorted 3DGS with an average L1 error of less than $10^{-4}$.