Research map / 3D vision / Gaussian splatting
FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting
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.
FlexGS trains a single 3D Gaussian Splatting model that, given a target ratio, learns to select a subset of Gaussians and transform them, enabling elastic-size rendering without per-ratio fine-tuning.
Problem
3DGS requires large GPU memory, and existing compression/pruning methods need separate runs or fine-tuning for each target size and device budget.
Approach
A Gumbel-Softmax-trained adaptive Gaussian selector conditioned on the input ratio, guided by a Global Importance score, combined with a ratio-conditioned Gaussian transform field (4D neural field with multi-head MLP predictor) that displaces selected Gaussians to compensate for removed ones.
Key results
- At 1% Gaussians on Mip-NeRF360, FlexGS reaches 22.730 PSNR vs 21.766 for C3DGS and 14.455 for LightGS*.
- On T&T at 1% ratio FlexGS attains 23.090 PSNR, vs 21.535 for C3DGS and 14.653 for LightGS*.
- On unseen ratios (2-8%) on Zip-NeRF, FlexGS outperforms all baselines by at least 0.5 PSNR, with elastic inference taking 0.13 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlexGS instead of LightGS | 3DGS · Mip-NeRF360, 1% | PSNR: 22.265 -> 22.730 | Helped |
| Use FlexGS instead of LightGS | 3DGS · T&T, 1% | PSNR: 22.795 -> 23.090 | Helped |
| Use FlexGS instead of C3DGS | 3DGS · Zip-NeRF, 1% | PSNR: 19.955 -> 20.963 | Helped |
| Add the Adaptive Selector (ablation) | FlexGS · T&T, 1% | PSNR: 14.653 -> 23.090 | Helped |
| Add the Adaptive Selector (ablation) | FlexGS · T&T, 5% | PSNR: 19.545 -> 26.189 | Helped |
| Add the Transform Field (ablation) | FlexGS · T&T, 1% | PSNR: 15.117 -> 23.090 | Helped |
| Add the Transform Field (ablation) | FlexGS · T&T, 10% | PSNR: 14.576 -> 27.069 | Helped |
| Add the sparsity constraint (ablation) | FlexGS · T&T, 1% | PSNR: 23.017 -> 23.090 | Helped |
| Add the sparsity constraint (ablation) | FlexGS · T&T, 15% | PSNR: 27.147 -> 27.302 | Hurt |
| Add Global Importance guidance (ablation) | FlexGS · T&T, 1% | PSNR: 19.472 -> 23.090 | Helped |
| Add Global Importance guidance (ablation) | FlexGS · T&T, 15% | PSNR: 24.642 -> 27.302 | Helped |
| Add sparsity constraint and Global Importance guidance (ablation) | FlexGS · T&T, 1% | PSNR: 18.363 -> 23.090 | Helped |
Surprising findings
- Removing the Transform Field causes severe degradation (PSNR ~14.6-15.1 across ratios on T&T), showing masking alone is insufficient.
Limitations
- Selection quality at very low ratios depends on Global Importance guidance; GsNet alone struggles to converge at high ratios.
- Evaluated only on static scenes from three datasets.
Experimental setup
Datasets: Mip-NeRF360 Zip-NeRF Tanks and Temples
Benchmarks: Mip-NeRF360 Zip-NeRF Tanks and Temples
Metrics: PSNR SSIM LPIPS inference time
Baselines: LightGaussian LightGaussian without fine-tuning (LightGS*) C3DGS EAGLES
Builds on: 3D Gaussian Splatting LightGaussian Gumbel-Softmax Matryoshka representation learning K-Planes
Closest papers by meaning
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.93)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.93)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
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.