Research map / 3D vision / Gaussian splatting
Regularization and primitive design for 3DGS
Improves splatting quality through regularization, new kernels and rendering optimizations.
213 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2024–2026) · an approach in Gaussian splatting, keyword cluster “3d gaussian · gaussian splatting · splatting 3dgs”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 4 of 6 tracked benchmarks.
31 papers in 2023–24 and 182 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 25.4% to 25.3%.
Open questions the data points to
- Mip-NeRF360 (PSNR): best went from 29.34 (up to 2024) to 30.13 in 2025 — 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · 14 citations)
- Tanks & Temples (PSNR): best went from 24.22 (up to 2024) to 25.08 in 2025 — 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · 14 citations)
- Mip-NeRF360 (SSIM): best went from 0.869 (up to 2024) to 0.893 in 2025 — 3D Student Splatting and Scooping (CVPR 2025 · 5 citations)
- Tanks & Temples (LPIPS): best went from 0.174 (up to 2024) to 0.138 in 2025 — 3D Student Splatting and Scooping (CVPR 2025 · 5 citations)
What looks settled
- Deep Blending (PSNR): best up to 2024 30.39, best in 2025–26 30.37 from 16 papers, at most 0.2% better, if at all.
- Mip-NeRF360 (LPIPS): best up to 2024 0.12, best in 2025–26 0.145 from 10 papers, at most 0.2% better, if at all.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2024 · Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · 257 citations)
- 2025 · PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · 47 citations)
- 2026 · PCGS: Progressive Compression of 3D Gaussian Splatting (AAAI 2026 · 9 citations)
Earliest papers here
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · 257 citations)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · 215 citations)
- GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (CVPR 2024 · 214 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
Mip-NeRF360 · PSNR (higher is better)
26 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 29.34 | Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering (NeurIPS 2024 · 4 citations) |
| 2025 | 30.13 | 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · 14 citations) |
Deep Blending · PSNR (higher is better)
20 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 30.39 | ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · 28 citations) |
Tanks & Temples · PSNR (higher is better)
17 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 24.22 | Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering (NeurIPS 2024 · 4 citations) |
| 2025 | 25.08 | 3D-HGS: 3D Half-Gaussian Splatting (CVPR 2025 · 14 citations) |
Mip-NeRF360 · LPIPS (lower is better)
16 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.12 | GaussianMarker: Uncertainty-Aware Copyright Protection of 3D Gaussian Splatting (NeurIPS 2024 · 16 citations) |
Mip-NeRF360 · SSIM (higher is better)
13 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.869 | Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering (NeurIPS 2024 · 4 citations) |
| 2025 | 0.893 | 3D Student Splatting and Scooping (CVPR 2025 · 5 citations) |
Tanks & Temples · LPIPS (lower is better)
9 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.174 | Normal-GS: 3D Gaussian Splatting with Normal-Involved Rendering (NeurIPS 2024 · 4 citations) |
| 2025 | 0.138 | 3D Student Splatting and Scooping (CVPR 2025 · 5 citations) |
What has been tried
107 papers here report 1,094 measured changes (a component added, removed or swapped, compared with a baseline): helped 75%, no clear effect 5%, hurt 18%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add learnable volume masking | Playroom · PSNR | 29.87 -> 29.91 | Helped | Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · 257 citations) |
| add the convolutional speed-up module | Replica · PSNR | 36.915 (±0.05) -> 37.012 (±0.07) | Helped | Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · 215 citations) |
| Add Hierarchical Gaussian Splatting | Grass-on-fire editing · editing control | Helped | GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (CVPR 2024 · 214 citations) | |
| initialize Gaussian positions with Point-E-generated geometry instead of an origin-centered Gaussian distribution | Text-to-3D generation, especially asymmetric prompts · geometry consistency | Helped | Text-to-3D using Gaussian Splatting (CVPR 2024 · 171 citations) | |
| add progressive growing | Tanks and Temples · PSNR | 22.57 -> 31.28 | Helped | COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · 142 citations) |
| Add multi-scale training | 360 dataset, 4x · PSNR | 22.50 -> 23.50 | Helped | Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering (CVPR 2024 · 102 citations) |
| Add Gaussian RoI gradient control | Doll, “Turn its mouth into red” · editing precision | Helped | GaussianEditor: Editing 3D Gaussians Delicately with Text Instructions (CVPR 2024 · 89 citations) | |
| add opacity and scale regularizers | MipNeRF 360 average · PSNR | 23.90 -> 29.72 | Helped | 3D Gaussian Splatting as Markov Chain Monte Carlo (NeurIPS 2024 · 54 citations) |
| add shape reset | Mip-NeRF360 · PSNR | 26.57 -> 26.91 | Helped | GES : Generalized Exponential Splatting for Efficient Radiance Field Rendering (CVPR 2024 · 53 citations) |
| Split pruning into two rounds with refinement after each, rather than one round | Mip-NeRF 360 · PSNR | 26.12 -> 26.67 | Helped | PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting (CVPR 2025 · 47 citations) |
| Add pipelined rendering | Truck 4K · average FPS | 190.8 -> 289.0 | Helped | FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · 46 citations) |
| add 3D Gaussian consensus | Rubble (Mill19) · PSNR | 22.80 -> 25.78 | Helped | DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · 39 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| 3DGS | Mip-NeRF 360 | PSNR | 27.21 – 29.04 | 11 papers, e.g. CVPR 2025 and CVPR 2025 |
| 3DGS | Deep Blending | PSNR | 28.04 – 29.69 | 9 papers, e.g. CVPR 2025 and CVPR 2025 |
| 3DGS | Mip-NeRF 360 | LPIPS | 0.182 – 0.24 | 6 papers, e.g. NeurIPS 2024 and CVPR 2025 |
| 3DGS | Mip-NeRF360 | FPS | 64.07 – 347 | 4 papers, e.g. CVPR 2025 and CVPR 2025 |
| 3DGS | Mip-NeRF 360 | SSIM | 0.803 – 0.872 | 4 papers, e.g. CVPR 2025 and CVPR 2025 |
| Scaffold-GS | Mip-NeRF360 | PSNR | 27.5 – 29.02 | 4 papers, e.g. CVPR 2025 and ICLR 2025 |
What reviewers will expect
Datasets and baselines named in the 96 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Mip-NeRF 360 · 34 of 96 papers
- Deep Blending · 19 of 96 papers
- Tanks & Temples · 16 of 96 papers
- Tanks and Temples · 10 of 96 papers
- DTU · 9 of 96 papers
- LLFF · 7 of 96 papers
- Replica · 6 of 96 papers
- Blender · 5 of 96 papers
Baselines
- 3DGS · 41 of 96 papers
- 2DGS · 13 of 96 papers
- Scaffold-GS · 12 of 96 papers
- Mip-NeRF360 · 9 of 96 papers
- GOF · 8 of 96 papers
- Mip-NeRF · 8 of 96 papers
- Mip-Splatting · 7 of 96 papers
- LightGaussian · 7 of 96 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 213 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 12 papers
- Peking University · 9 papers
- Nanyang Technological University · 8 papers
- The University of Hong Kong · 7 papers
- The Chinese University of Hong Kong · 7 papers
- NVIDIA · 6 papers
- Zhejiang University · 6 papers
- National University of Singapore · 5 papers
- Huawei · 4 papers
- Hong Kong University of Science and Technology · 4 papers
Use it
Download BibTeX of 20 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Compact and efficient Gaussian splatting (269 papers)
- Sparse-view Gaussian splatting for novel views (111 papers)
- Animatable Gaussian human avatars (42 papers)
Elsewhere in the topic
- Deformable 4D Gaussian splatting (112 papers)
- Monocular dynamic scene reconstruction (80 papers)
- Animatable Gaussian human avatars (14 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).