Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields

TL;DR

Turbo-GS addresses the slow fitting of 3D Gaussian Splatting by introducing a color-appearance guided densification strategy, a convergence-aware budget for Gaussian growth, and selective densification that focuses on frequently visited regions. It also introduces a dilation-based rendering pipeline to enable efficient high-resolution (4K) training while preserving rendering quality. Built on Scaffold-GS, Turbo-GS combines better initialization, batched training, and efficient rendering to achieve multi-fold speedups with competitive or superior view synthesis quality. Extensive experiments on MipNeRF-360 and Deep Blending demonstrate faster optimization without sacrificing perceptual fidelity, making high-quality radiance fields more practical for large-scale and high-resolution applications.

Abstract

Novel-view synthesis is an important problem in computer vision with applications in 3D reconstruction, mixed reality, and robotics. Recent methods like 3D Gaussian Splatting (3DGS) have become the preferred method for this task, providing high-quality novel views in real time. However, the training time of a 3DGS model is slow, often taking 30 minutes for a scene with 200 views. In contrast, our goal is to reduce the optimization time by training for fewer steps while maintaining high rendering quality. Specifically, we combine the guidance from both the position error and the appearance error to achieve a more effective densification. To balance the rate between adding new Gaussians and fitting old Gaussians, we develop a convergence-aware budget control mechanism. Moreover, to make the densification process more reliable, we selectively add new Gaussians from mostly visited regions. With these designs, we reduce the Gaussian optimization steps to one-third of the previous approach while achieving a comparable or even better novel view rendering quality. To further facilitate the rapid fitting of 4K resolution images, we introduce a dilation-based rendering technique. Our method, Turbo-GS, speeds up optimization for typical scenes and scales well to high-resolution (4K) scenarios on standard datasets. Through extensive experiments, we show that our method is significantly faster in optimization than other methods while retaining quality. Project page: https://ivl.cs.brown.edu/research/turbo-gs.

Figures (9)

• Figure 1: Turbo-GS accelerates 3DGS fitting significantly while preserving rendering quality. It proposes efficient densification strategy and innovative dilated rendering allow training on 4K images in minutes—significantly outperforming baseline methods. Notably, Turbo-GS converges on the 4K bicycle scene in just 13 minutes—over 3$\times$faster than Taming 3DGS (40 minutes), 14$\times$ faster than 3DGS (187 minutes) and Scaffold-GS (185 minutes).
• Figure 2: Effect of Densification Rate. This plot shows the effect of densification rate with Scaffold-GS scaffoldgs versus Turbo-GS (Ours) on the Bicycle scene mip360. Scaffold-GS with densification every 100 iterations (default, orange) takes time to converge. An aggressive version of Scaffold-GS with densification every 20 iterations (green) initially shows improved convergence, but plateaus afterward. Ours (blue) produces higher-quality reconstruction with densification every 20 iterations.
• Figure 3: Different Gradients Schema. We find that covariance-based densification yields subpar quality; and in practice, opacity-based densification always leads to floaters. Color and position gradients are more reliable.
• Figure 4: Gradient Visualization. We rasterize the Gaussian gradient into image plane and observe that: (a) Position Gradients focus only on certain regions in the scene, while (b) Color Gradients provide cues from overall regions. These are useful for regions such as grass and background structure.
• Figure 5: Loss analysis with power function fitting. For all scenes, the $log$(loss) is linear to the $log$(iterations) after the initial stage. Thus, the relation between iteration and convergence follows a power function. We design a power-law-based adaptive budget schedule based on these insights.