Research map / 3D vision / Gaussian splatting
GES : Generalized Exponential Splatting for Efficient Radiance Field Rendering
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 GES, a 3D radiance field representation that replaces Gaussian splats with generalized exponential functions with a learnable shape parameter β, plus a frequency-modulated image loss, to reduce memory and increase rendering speed.
Problem
3D Gaussian Splatting needs very many Gaussians to represent sharp edges and high-frequency content because Gaussians are low-pass, causing large memory footprints for novel view synthesis.
Approach
Replace each Gaussian splat with a generalized exponential function that has a learnable shape exponent β (β=2 recovers a Gaussian), rasterized approximately by rescaling covariance with a Gamma-function factor, and trained with a frequency-modulated edge-aware loss that moves from low to high frequencies.
Key results
- On Mip-NeRF360, GES reaches 26.91 PSNR and 377MB memory versus 3D Gaussians-30K's 27.21 PSNR and 734MB.
- GES achieves up to 39% higher rendering speed (e.g., 186 FPS vs 134 FPS on Mip-NeRF360).
- On Deep Blending, GES reaches 29.68 PSNR with 399MB versus 29.41 PSNR and 676MB for 3D Gaussians-30K.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Gaussian splats with generalized exponential splats | GES · Mip-NeRF360 | FPS: 134 -> 186 | Helped |
| replace Gaussian splats with generalized exponential splats | GES · Mip-NeRF360 | Mem: 734MB -> 377MB | Helped |
| replace Gaussian splats with generalized exponential splats | GES · Tanks&Temples | PSNR: 23.14 -> 23.35 | Helped |
| replace Gaussian splats with generalized exponential splats | GES · Deep Blending | PSNR: 29.41 -> 29.68 | Helped |
| add shape reset (ablation) | GES · Mip-NeRF360 | PSNR: 26.57 -> 26.91 | Helped |
| add frequency-modulated loss Lω (ablation) | GES · Mip-NeRF360 | Size (MB): 411 -> 377 | Helped |
| add frequency-modulated loss Lω (ablation) | GES · Mip-NeRF360 | PSNR: 27.07 -> 26.91 | Hurt |
| add frequency-modulated loss Lω (ablation) | GES · Mip-NeRF360 | SSIM: 0.800 -> 0.794 | Hurt |
Surprising findings
- Adding the frequency-modulated loss reduces memory (411MB to 377MB) with nearly unchanged PSNR in the ablation.
Limitations
- Pushing compactness lowers PSNR, especially on complex scenes, due to pruning tied to beta tuning.
- View-dependent boundary error from the approximate rasterization via covariance rescaling.
- Not evaluated on dynamic or more complex environments.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS training time FPS memory size
Baselines: 3D Gaussians-30K 3D Gaussians-7K Instant-NGP Mip-NeRF360 Plenoxels
Builds on: 3D Gaussian Splatting Zwicker EWA splatting BARF frequency modulation Mip-NeRF360
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.90)
- DiffGS: Functional Gaussian Splatting Diffusion (NeurIPS 2024 · similarity 0.90)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · similarity 0.89)
- DoF-Gaussian: Controllable Depth-of-Field for 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
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.