FastGS: Training 3D Gaussian Splatting in 100 Seconds

2025-11-06

AI Summary

To address the inefficiency and redundant computation in 3D Gaussian Splatting (3DGS) training caused by static, pre-defined Gaussian budgets, this paper proposes FastGS—a fast training framework guided by multi-view consistency. Instead of relying on fixed budget constraints, FastGS employs multi-view consistency as a unified, unsupervised signal to dynamically drive adaptive densification and pruning of Gaussians during training, enabling geometry-aware and rendering-fidelity-preserving point-cloud optimization. This approach significantly improves training throughput without compromising visual quality. On the Mip-NeRF 360 and Deep Blending benchmarks, FastGS achieves speedups of 3.32× and 15.45×, respectively; in cross-dataset generalization tests, it consistently delivers 2–7× acceleration while maintaining rendering quality comparable to the original 3DGS method.

Technology Category

• Computer Vision: 3D Computer Vision
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Efficient ML / Green AI

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

The dominant 3D Gaussian splatting (3DGS) acceleration methods fail to properly regulate the number of Gaussians during training, causing redundant computational time overhead. In this paper, we propose FastGS, a novel, simple, and general acceleration framework that fully considers the importance of each Gaussian based on multi-view consistency, efficiently solving the trade-off between training time and rendering quality. We innovatively design a densification and pruning strategy based on multi-view consistency, dispensing with the budgeting mechanism. Extensive experiments on Mip-NeRF 360, Tanks&Temples, and Deep Blending datasets demonstrate that our method significantly outperforms the state-of-the-art methods in training speed, achieving a 3.32$ imes$ training acceleration and comparable rendering quality compared with DashGaussian on the Mip-NeRF 360 dataset and a 15.45$ imes$ acceleration compared with vanilla 3DGS on the Deep Blending dataset. We demonstrate that FastGS exhibits strong generality, delivering 2-7$ imes$ training acceleration across various tasks, including dynamic scene reconstruction, surface reconstruction, sparse-view reconstruction, large-scale reconstruction, and simultaneous localization and mapping. The project page is available at https://fastgs.github.io/

Problem

• Regulating Gaussian numbers in 3DGS training to reduce computational overhead
• Solving training time versus rendering quality trade-off in 3D reconstruction
• Developing efficient acceleration framework for various 3D scene reconstruction tasks

Innovation

• Uses multi-view consistency for Gaussian importance evaluation
• Implements densification and pruning without budgeting mechanism
• Achieves 3-15x acceleration across diverse reconstruction tasks