Research map / 3D vision / Gaussian splatting
FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution 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.
Introduces FlashGS, a CUDA rasterization library for 3D Gaussian Splatting that removes redundant Gaussian-tile pairs via precise ellipse-tile intersection, balances work by Gaussian size, and pipelines color computation to render large-scale, high-resolution scenes in real time.
Problem
Real-time 3DGS rendering is too slow for large-scale, high-resolution scenes because the tile-based rasterizer produces many redundant Gaussian-tile pairs and imbalanced workloads.
Approach
Replace the bounding-box overlap test with an exact opacity-aware ellipse-tile intersection computed via algebraic simplifications, assign threads adaptively by Gaussian size, and software-pipeline the feature loads and alpha blending, all without changing the scene representation.
Key results
- FlashGS-Full achieves an average 7.2× speedup over vanilla 3DGS across scenes and 8.7× on the 7 large-scale or high-resolution tests.
- FlashGS-Full reaches 1.8× average speedup over AdR-Gaussian (AdRGS-Full), and 2.2× on large-scale/high-resolution scenes.
- FlashGS sustains at least 125.9 FPS for the slowest frame on all tested scenes, with image quality close to vanilla 3DGS (e.g. Truck PSNR 27.20 vs 27.24).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlashGS-Full instead of AdRGS-Full | 3DGS rendering pipeline · Truck 4K | average FPS: 166.2 -> 289.0 | Helped |
| Use FlashGS-Full instead of AdRGS-Full | 3DGS rendering pipeline · Train 4K | average FPS: 176.5 -> 301.2 | Helped |
| Use FlashGS-Full instead of AdRGS-Full | 3DGS rendering pipeline · Playroom 4K | average FPS: 211.0 -> 367.7 | Helped |
| Use FlashGS-Full instead of AdRGS-Full | 3DGS rendering pipeline · DrJohnson 4K | average FPS: 196.1 -> 334.4 | Helped |
| Use FlashGS-Full instead of AdRGS-AABB | 3DGS rendering pipeline · MatrixCity 4K | average FPS: 68.8 -> 207.4 | Helped |
| Use FlashGS-Full instead of AdRGS-AABB | 3DGS rendering pipeline · Rubble 4608*3456 | average FPS: 59.0 -> 183.9 | Helped |
| Add pipelined rendering (ablation) | FlashGS · Truck 4K | average FPS: 190.8 -> 289.0 | Helped |
| Add pipelined rendering (ablation) | FlashGS · Train 4K | average FPS: 203.8 -> 301.2 | Helped |
| Add pipelined rendering (ablation) | FlashGS · Playroom 4K | average FPS: 223.6 -> 367.7 | Helped |
| Add pipelined rendering (ablation) | FlashGS · DrJohnson 4K | average FPS: 204.8 -> 334.4 | Helped |
| Add pipelined rendering (ablation) | FlashGS · MatrixCity 4K | average FPS: 117.1 -> 207.4 | Helped |
| Add pipelined rendering (ablation) | FlashGS · Rubble 4608*3456 | average FPS: 97.6 -> 183.9 | Helped |
Surprising findings
- Hardware rasterization pipelines (web-splat, fast-gaussian-rasterization) are much slower than the optimized CUDA path, attributed to quad redundancy and small-triangle inefficiency.
- AdR-Gaussian retrained model shows noticeably lower image quality (e.g. Truck PSNR 26.83 vs 27.24) than vanilla 3DGS.
Limitations
- Gains are reported on a single GPU (RTX 3090) only.
- On low-resolution Truck-1080p and Train-1080p, AdR-Gaussian has higher slowest-frame FPS than FlashGS.
- Adaptive scheduling is needed; without it (FlashGS-w/o) performance drops below AdR-AABB.
Experimental setup
Datasets: Tanks and Temples (Truck, Train) Deep Blending (Playroom, DrJohnson) MatrixCity Rubble
Benchmarks: Truck Train Playroom DrJohnson MatrixCity Rubble
Metrics: FPS (average) FPS (slowest frame, MinFPS) PSNR SSIM LPIPS max Gaussian-tile intersections
Baselines: 3DGS gsplat AdR-Gaussian web-splat fast-gaussian-rasterization
Builds on: 3D Gaussian Splatting AdR-Gaussian gsplat
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds (CVPR 2025 · similarity 0.93)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.93)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · 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.