Skip to content

DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds

Conference: CVPR 2025
arXiv: 2503.18402
Code: dashgaussian.github.io
Area: 3D Vision / 3DGS Acceleration
Keywords: 3D Gaussians, Training Acceleration, Frequency Scheduling, Resolution Scheduling, Adaptive Densification

TL;DR

DashGaussian is proposed, offering a joint framework for scheduling rendering resolution and Gaussian primitive count based on frequency analysis. It reformulates 3DGS optimization as a progressive fitting of high-frequency components, achieving an average acceleration of 45.7% without compromising rendering quality.

Background & Motivation

  • Although 3DGS is significantly faster than NeRF (tens of minutes vs. days), further acceleration is still required on resource-constrained devices and for large-scale scene reconstruction.
  • Existing acceleration methods fall into two categories: (1) engineering optimizations (e.g., efficient forward/backward implementations); (2) algorithmic optimizations (e.g., pruning redundant Gaussians), where the latter often sacrifices rendering quality.
  • Three key observations: a) computational overhead is primarily determined by rendering resolution and Gaussian count; b) at the beginning of optimization, Gaussians are sparse, making high-resolution rendering wasteful; c) in the late stages of optimization, a surge in Gaussian count brings massive computation with limited quality improvement.
  • Key Insight: Allocating computational resources reasonably is more efficient than simply pruning parameters.

Method

Overall Architecture

The optimization of 3DGS is reformulated as a progressive process of fitting higher-frequency components in training views. Based on this, a resolution scheduler and a primitive scheduler are proposed to jointly control the optimization complexity.

Key Designs

  1. Frequency-Guided Resolution Scheduler:

    • Function: Adaptively increases the rendering resolution step-by-step during optimization.
    • Mechanism: Formulating image downsampling as removing high-frequency components, a resolution saliency function is defined as \(\mathcal{X}(\mathbf{F}) = \frac{1}{N}\sum_{n=1}^N \sum_{i,j} \|\mathbf{F}^n(i,j)\|_2\). Iteration steps for high/low resolutions are allocated using the frequency energy ratio \(f(\mathbf{F}, \mathbf{F}_r) = \frac{\mathcal{X}(\mathbf{F}) - \mathcal{X}(\mathbf{F}_r)}{\mathcal{X}(\mathbf{F})}\).
    • Design Motivation: Low-frequency components contain core structural information and require less computation, so they should be fitted first. High-frequency details require more primitives to be effectively fitted and should be addressed after sufficient primitives are present.
    • Switching Point: Switches from low resolution to high resolution at the \(s_r = S \cdot \mathcal{X}(\mathbf{F}_r) / \mathcal{X}(\mathbf{F})\) iteration.

  2. Resolution-Guided Primitive Scheduler:

    • Function: Controls the growth of Gaussian primitives in synchronization with the resolution.
    • Mechanism: \(P_i = P_{\text{init}} + (P_{\text{fin}} - P_{\text{init}}) / (r^{(i)})^{2-i/S}\), where the power factor decays linearly from 2 to 1, producing a concave growth curve (suppressed in the early stage, encouraged in the middle stage).
    • Design Motivation: A specific resolution should match an appropriate number of primitives. Too many primitives in low-resolution stages are wasteful and may lead to over-densification. The concave curve ensures primitives grow rapidly only when truly needed.

  3. Momentum-Based Primitive Budget Estimation:

    • Function: Adaptively determines the final primitive count \(P_{\text{fin}}\) without relying on dataset priors.
    • Mechanism: Treating \(P_{\text{fin}}\) as momentum and the natural densification amount at each step \(P_{\text{add}}\) as force: \(P_{\text{fin}}^{(i)} = \max(P_{\text{fin}}^{(i-1)}, \gamma P_{\text{fin}}^{(i-1)} + \eta P_{\text{add}}^{(i)})\).
    • Design Motivation: Bypasses artificially set upper bounds, dynamically adjusting the primitive budget based on actual scene requirements.

Loss & Training

  • Uses anti-aliasing downsampling (frequency domain center cropping + DFT/IDFT) to avoid 2D aliasing.
  • Discretizes the resolution factor using floor functions to encourage primitives to grow faster than under continuous scheduling.
  • Hyperparameters: \(a=4\) (maximum downsampling corresponding to \(1/4\) of frequency energy), \(\gamma=0.98\), \(\eta=1\), \(P_{\text{fin}}^{(0)} = 5 \cdot P_{\text{init}}\).
  • Maintains the original hyperparameters unchanged when incorporating various 3DGS backbones.

Key Experimental Results

Main Results (Comparison with Fast Optimization Methods)

Method Mip-NeRF360 PSNR Time (min) Deep Blending PSNR Time (min) T&T PSNR Time (min)
3DGS 27.72 18.31 29.50 17.27 23.62 10.59
Reduced-3DGS 27.28 15.52 29.78 13.74 23.59 8.31
Taming-3DGS 27.61 5.51 29.69 4.52 23.62 4.02
DashGaussian 27.92 3.23 30.02 2.20 23.97 2.62

Speedup Effects on Various Backbones

Backbone Original Time +DashGaussian Speedup Ratio PSNR Change
3DGS 18.31 min 10.16 min 44.5% +0.09
Mip-Splatting 25.83 min 12.60 min 51.2% +0.08
Revising-3DGS 5.73 min 3.46 min 39.6% +0.20
Taming-3DGS 5.51 min 3.23 min 41.4% +0.31

Ablation Study

Configuration Key Metrics Description
Resolution scheduling only Significant speedup but slight quality drop Lack of primitive coordination leads to underfitting
Primitive scheduling only Moderate speedup High-resolution rendering remains the bottleneck
Joint scheduling (Full) Maximum speedup with maintained quality Optimal coordination between resolution and primitives
Fixed \(P_{\text{fin}}\) Over/under-densification in some scenes Momentum estimation yields better adaptivity

Key Findings

  • Accelerates optimization by 45.7% on average with widespread improvements in PSNR (speed is not gained at the cost of quality).
  • Integrates seamlessly with arbitrary 3DGS backbones as a plug-and-play accelerator.
  • The concave primitive growth curve (fewer in early stages, more in late stages) outperforms existing convex or front-heavy strategies.
  • Compresses the optimization time of Taming-3DGS on the Mip-NeRF 360 dataset to approximately 3 minutes (~200 seconds level).

Highlights & Insights

  • Unifies the theoretical foundation of resolution scheduling from the perspective of frequency analysis, reformulating 3DGS optimization as a progressive frequency fitting process.
  • The philosophy of "allocating computational resources reasonably rather than pruning parameters" is more elegant than aggressive pruning and avoids quality loss.
  • Momentum-based primitive budgeting is a truly adaptive solution, escaping the limitations of manually defined upper bounds.
  • Experiments demonstrate that reasonable allocation of computational resources can even enhance quality (the low-resolution stage acts as implicit regularization).

Limitations & Future Work

  • Discretizing the resolution factor (using floor) is an engineering approximation, leaving the theoretical optimality of continuous scheduling not fully realized.
  • The frequency energy ratio \(f\) assumes perfect anti-aliased downsampling, whereas residual aliasing may exist in practice.
  • The selection of hyperparameter \(a=4\) is not extensively ablated.
  • Not fully validated on ultra-large-scale scenes (e.g., city-scale).
  • 3DGS \(\to\) Base representation; Mip-Splatting \(\to\) Multi-scale rendering and anti-aliasing.
  • Taming-3DGS \(\to\) Engineering acceleration baseline; Mini-Splatting \(\to\) Algorithmic pruning baseline.
  • Nyquist Shannon sampling theorem + DFT \(\to\) Theoretical foundation for resolution scheduling.

Rating

  • Novelty: ⭐⭐⭐⭐ Frequency-guided resolution scheduling offers a novel perspective, and the joint scheduling scheme is designed cleverly.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ High thoroughness with three datasets, four backbones, and detailed ablations/comparisons.
  • Writing Quality: ⭐⭐⭐⭐ Clear theoretical derivations; the frequency analysis perspective is highly educational.
  • Value: ⭐⭐⭐⭐⭐ Extremely high practical value as a plug-and-play general accelerator.