Source-linked AI summary
FastGS: Training 3D Gaussian Splatting in 100 Seconds
Shiwei Ren, Tianci Wen, Yongchun Fang, Biao Lu
TL;DR
3DGS accelerates training relative to NeRF, but redundant Gaussians and inefficient density control remain obstacles to practical training speed. FastGS guides densification and pruning by multi-view reconstruction quality, and reports faster training while maintaining comparable rendering quality across static scenes and a range of tasks.
Problem
3DGS training can take tens of minutes per scene, while existing density-control strategies may retain redundant Gaussians and inadequately balance representation size with rendering quality.
Source: §1 · p. 1, §4.2 · p. 4Method
FastGS uses multi-view reconstruction quality to guide Gaussian densification and pruning, avoiding a budgeting mechanism.
Source: §1 · p. 2Results
FastGS reports the fastest training among state-of-the-art methods with comparable rendering quality on static scenes, and reduced training time across various tasks.
Source: §6 · p. 9Takeaways & Limitations
The authors report that FastGS generalizes across tasks, including dynamic scene reconstruction, surface reconstruction, sparse-view reconstruction, large-scale reconstruction, and SLAM.
Source: §6 · p. 9FastGS works best when training from sparse point clouds; post-training dense feed-forward 3DGS outputs challenges pruning, with 3K iterations taking about 20 seconds.
Source: §5.5 · pp. 8–9
Abstract
from arXiv · showhide
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$\times$ training acceleration and comparable rendering quality compared with DashGaussian on the Mip-NeRF 360 dataset and a 15.45$\times$ acceleration compared with vanilla 3DGS on the Deep Blending dataset. We demonstrate that FastGS exhibits strong generality, delivering 2-7$\times$ 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/
1. Introduction
3DGS speeds training and rendering relative to NeRF, but scene training still takes tens of minutes, and redundant Gaussians and rendering inefficiencies remain bottlenecks. FastGS uses multi-view reconstruction quality to guide densification and pruning without a budget mechanism.
2. Related Work
1. Introduction
- FastGS uses multi-view reconstruction quality to guide Gaussian densification and pruning, aiming to reduce redundancy without a budget mechanism while maintaining comparable rendering quality.The introduction says it can train a scene in around 100 seconds.Source: §1 · p. 2
3. Background: 3D Gaussian Splatting
- 3DGS represents scenes with anisotropic 3D Gaussians initialized from sparse SfM points, which are projected into 2D image-plane Gaussians for rendering.Each primitive has a mean, rotation, scale, opacity, and color coefficients; projected Gaussians contribute to pixels through alpha blending.Source: §3 · p. 3
4. FastGS
FastGS uses multi-view reconstruction quality to guide Gaussian densification and pruning, and adds compact-box rasterization to reduce unnecessary Gaussian–tile pairs. It retains vanilla 3DGS’s L1-plus-SSIM optimization objective.
4.1. Overview
- FastGS scores Gaussian importance from multi-view reconstruction quality and uses that information to guide both densification and pruning.Unlike Taming-3DGS and Speedy-Splat, it does not rely on Gaussian-associated scores for these controls.Source: §4.1 · p. 3
4.2. Multi-view Consistent Densification
- VCD averages high-error pixels within each Gaussian’s projected footprint across sampled views to identify under-reconstructed regions for densification.High-error pixels are selected from normalized per-pixel L1 loss maps, and a score threshold determines densification candidates.Source: §4.2 · pp. 4–5
4.3. Multi-view Consistent Pruning
- VCP scores each Gaussian by its impact on reconstruction quality across views and prunes those whose scores exceed a threshold.The score incorporates photometric loss, including mean absolute error and structural similarity, to identify Gaussians with low contribution.Source: §4.3 · p. 5
4.4. Compact Box
- Compact box extends precise tile intersection by pruning Gaussian–tile pairs with minimal contribution based on distance from the Gaussian center.It reduces unnecessary pairs while maintaining quality.Source: §4.4 · pp. 5–6, Figure 4 · p. 5
4.5. Optimization
- FastGS optimizes learnable parameters using rendered-pixel L1 loss combined with an SSIM term, as in vanilla 3DGS.Source: §4.5 · p. 6
5. Experiments
Experiments compare FastGS with accelerated 3DGS methods, test its application across reconstruction tasks and backbones, and examine its component contributions and limitations.
5.1. Experimental Setup
- Tests span three real-world datasets and dynamic, surface, sparse-view, large-scale, and SLAM tasks, evaluating image quality, training time, Gaussian count, and rendering speed.Methods use 30K iterations and an NVIDIA RTX 4090 GPU.Source: §5.1 · p. 6
5.2. Comparison with Fast Optimization Methods
- FastGS trains a scene in around 100 seconds, with a fastest case of 77 seconds, while offering comparable rendering quality to other methods.Source: §5.2 · p. 7, Table 1 · p. 7
- FastGS-Big exceeds DashGaussian by more than 0.2 dB in rendering quality, reduces training time by 43.6%, and cuts Gaussian count by more than half.Source: §5.2 · p. 8
5.3. Generality of FastGS
- FastGS accelerates baselines by 2–6× across varied tasks and achieves 3–8× faster training across backbones while maintaining rendering quality.Source: §5.3 · p. 8, Table 6 · p. 8
5.4. Ablation Study
- VCD yields over 2× faster training and 80% fewer Gaussians; VCP cuts training time 25% and Gaussian count 26%, while CB cuts time 14%.VCD and VCP preserve rendering quality; CB achieves comparable quality.Source: §5.4 · p. 8, Table 7 · p. 8
5.5. Discussions and Limitations
- Dense outputs from feed-forward 3DGS make it difficult to prune many Gaussians within a few thousand iterations while maintaining quality; 3K post-training iterations take about 20 seconds.Source: §5.5 · pp. 8–9
6. Conclusion
FastGS uses multi-view-consistent densification and pruning to accelerate 3DGS training while retaining comparable rendering quality, with gains across tasks.
- FastGS is a general 3DGS acceleration framework using multi-view-consistent densification and pruning to prevent redundant Gaussians.Source: §6 · p. 9
- Experiments report the fastest training among state-of-the-art methods while maintaining comparable rendering quality.Source: §6 · p. 9
- Results show reduced training time across varied tasks, demonstrating FastGS’s generality.Source: §6 · p. 9
7. Overview
The supplementary material details the compact box, additional experiments, and scene-wise results.
- Section 8 describes the proposed compact box.Source: §7 · p. 12
- Section 9 provides additional experimental details, including implementation and integration across tasks and backbones.Source: §7 · p. 12
- Sections 10–12 cover computational overhead, additional ablations, and scene-wise results.Source: §7 · p. 12
8. Details of Compact Box
Compact Box (CB) refines tile intersection by discarding Gaussian–tile pairs whose pixel contributions are negligible. A Mahalanobis-distance threshold controls each 2D Gaussian’s effective support; a smaller β tightens that support and reduces rasterization work.
8. Details of Compact Box
- Because alpha decays exponentially with Mahalanobis distance, pairs involving pixels far from a Gaussian center can be discarded with minimal rendering effect.The method defines a Mahalanobis-distance threshold for pruning pairs with negligible contributions.Source: §8 · p. 12
- CB reduces Gaussian–tile pairs compared with vanilla 3DGS and Speedy-Splat by shrinking effective 2D Gaussian regions.The method refines Speedy-Splat’s precise tile-intersection strategy.Source: Figure 7 · p. 12, §8 · p. 12
- A smaller β tightens the ellipse around the 2D Gaussian mean, reducing its spatial extent, affected pixels, redundant pairs, and rasterization time.CB integrates this support adjustment into Speedy-Splat’s snugbox strategy.Source: §8 · p. 12
9. More Details
FastGS combines VCD, VCP, and CB with an accelerated 3DGS baseline, applying Gaussian-count control across several reconstruction tasks and backbones. The details describe its implementation choices and report a 2.70× average training speed-up for SLAM.
9.1. Implementation Details
- VCD and VCP keep Gaussian counts low throughout training and account for most of FastGS’s acceleration, while CB and absolute gradients contribute marginally to that aspect.The implementation combines these components with the 3DGS-accel baseline and uses absolute gradients from AbsGS for splitting.Source: §9.1 · pp. 12–13
9.2. Generalizing FastGS to Other Tasks
- FastGS applies VCD and VCP to dynamic, surface, sparse-view, and SLAM reconstruction, while large-scale reconstruction uses VCD without VCP because pruning is at the anchor level.CB is integrated into rasterization across these task adaptations.Source: §9.2 · p. 13
- 2.70× average training speed-up is reported for FastGS on SLAM.The SLAM setup integrates VCD, VCP, and CB into Photo-SLAM.Source: Table 9 · p. 13, §9.2 · p. 13
9.3. Equipping FastGS to Backbones
- FastGS equips Mip-Splatting and Scaffold-GS backbones with CB; VCD and VCP replace the former’s density control, while Scaffold-GS uses VCD alone because pruning is anchor-level.The table is described as reporting quantitative results for accelerating various backbones.Source: §9.3 · pp. 13–14, Table 10 · p. 14
10. Computational Overhead
The paper reports that FastGS uses relatively little GPU memory, making it suitable for devices with limited resources.
10. Computational Overhead
- FastGS requires relatively low GPU memory, making it suitable for resource-limited devices.Computational resource consumption is reported in Table 11.Source: §10 · p. 14
11. Additional Ablation
Ablations identify VCD and VCP as the largest contributors to acceleration, and examine how densification threshold and sampled-view count affect efficiency and rendering quality.
Component-wise Ablation
- In component-wise ablations, VCD and VCP achieved significantly greater speed-up than absolute gradients or CB, which contributed little to acceleration, by strictly controlling Gaussian counts and keeping them low throughout training.Source: §11 · p. 14
VCD Threshold τd
- On Tanks & Temples, a larger τd reduces Gaussian count and training time but slightly degrades quality; the default τd = 5 balances efficiency and fidelity.A smaller threshold allows more Gaussians to be densified and yields slightly higher quality at increased training time.Source: §11 · p. 14
Number of sampled views K
- On Mip-NeRF 360, too few sampled views slightly degrade quality, while more views increase training time with minimal improvement; the default K = 10 balances both.Source: §11 · p. 14
12. Scene-wise Results
The section presents scene-wise quantitative results for Mip-NeRF 360, Deep Blending, and Tanks & Temples, alongside qualitative comparisons across multiple scenes.
12. Scene-wise Results
- The supplementary tables include ablations on densification threshold and sampled-view count, plus comparisons of optimization and rasterization strategies.The sampled-view caption specifies that “all” means using all training views.Source: Table 13 · p. 15, Table 14 · p. 15, Table 15 · p. 15
- Scene-wise quantitative results are presented for Mip-NeRF 360, Deep Blending, and Tanks & Temples.The supplied table captions identify each dataset but do not provide numerical results.Source: Table 17 · p. 16, Table 18 · p. 16, Table 19 · p. 16
- Additional visual comparisons cover scenes from Mip-NeRF 360, Deep Blending, and other tasks.The captions name bicycle, garden, stump, room, counter, kitchen, bonsai, playroom, flame salmon, cut roasted beef, Caterpillar, Meetingroom, trex, leaves, amsterdam, and office0.Source: Figure 8 · p. 17, Figure 9 · p. 18, Figure 10 · p. 19