Research map / 3D vision / Gaussian splatting
FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FreGS, a 3D Gaussian splatting method that adds progressive (low-to-high) frequency regularization in Fourier space to fix over-reconstruction during Gaussian densification.
Problem
3D Gaussian splatting suffers from over-reconstruction during densification, where high-variance regions are covered by a few large Gaussians, causing blur and artifacts in novel view synthesis.
Approach
Regularize amplitude and phase discrepancies between rendered and ground-truth image spectra, annealing from low-frequency to high-frequency components via low-pass and dynamic high-pass Fourier filters, to guide coarse-to-fine Gaussian densification.
Key results
- On Mip-NeRF360, FreGS reaches 27.85 PSNR / 0.826 SSIM / 0.209 LPIPS vs 3D-GS 27.21 / 0.815 / 0.214.
- On Tanks&Temples, FreGS reaches 23.96 PSNR vs 3D-GS 23.14 PSNR.
- On Deep Blending, FreGS reaches 29.93 PSNR vs 3D-GS 29.41 PSNR; ablation shows frequency annealing adds gains over plain frequency regularization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add frequency regularization (FR) on amplitude and phase discrepancies (ablation) | 3D Gaussian splatting · Mip-NeRF360 | PSNR: +0.42 (27.21 to 27.63) | Helped |
| add frequency regularization (FR) on amplitude and phase discrepancies (ablation) | 3D Gaussian splatting · Tank&Temple | PSNR: +0.62 (23.14 to 23.76) | Helped |
| add frequency annealing (progressive low-to-high frequency regularization) (ablation) | Base+FR · Mip-NeRF360 | PSNR: +0.22 (27.63 to 27.85) | Helped |
| add frequency annealing (progressive low-to-high frequency regularization) (ablation) | Base+FR · Tank&Temple | PSNR: +0.20 (23.76 to 23.96) | Helped |
| progressive frequency regularization with annealing (FreGS) vs 3D-GS | 3D Gaussian splatting · Mip-NeRF360 | PSNR: +0.64 (27.21 to 27.85) | Helped |
| progressive frequency regularization with annealing (FreGS) vs 3D-GS | 3D Gaussian splatting · Tank&Temple | PSNR: +0.82 (23.14 to 23.96) | Helped |
| progressive frequency regularization with annealing (FreGS) vs 3D-GS | 3D Gaussian splatting · Deep Blending | PSNR: +0.52 (29.41 to 29.93) | Helped |
Surprising findings
- Naive frequency regularization without low-to-high annealing biases training toward artifacts, showing annealing is necessary.
Limitations
- Frequency regularization is terminated when densification ends, so its effect is confined to the densification phase.
- Gains are modest on some metrics (e.g., LPIPS on Deep Blending nearly unchanged vs 3D-GS).
- Hyperparameters such as the frequency schedule and loss weights are hand-set.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS
Baselines: 3D-GS Mip-NeRF360 InstantNGP Plenoxels
Builds on: 3D Gaussian splatting NeRF BARF
Closest papers by meaning
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.89)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.89)
- Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting (AAAI 2025 · similarity 0.89)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering (CVPR 2024 · similarity 0.89)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.