Research map / Point clouds / Shape and surface completion
Point cloud upsampling and normals
Upsample point clouds and estimate normals with learned surface refinement.
64 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Shape and surface completion, keyword cluster “reconstruction · implicit · meshes”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 6 tracked benchmarks.
19 papers in 2023–24 and 29 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 21.3% to 25.7%.
Open questions the data points to
- PU-GAN 4× (CD (10^-3)): best went from 0.131 (up to 2024) to 0.119 in 2025 — Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · 4 citations)
- PU1K 4× (CD (10^-3)): best went from 0.217 (up to 2024) to 0.155 in 2025 — Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · 4 citations)
- Papers report Grad-PU on PUGAN 16× (CD × 10−3) anywhere from 0.108 to 0.129: comparisons against it depend on whose number you take.
No longer reported by papers here: PU-GAN 4× (HD (10^-3), last in 2024); PU-GAN 4× (P2F (10^-3), last in 2024); PU-GAN 16× (HD (×10^-3), last in 2024); PU-GAN 16× (P2F (×10^-3), last in 2024).
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
- 2016 · Sorting out typicality with the inverse moment matrix SOS polynomial (NeurIPS 2016 · 31 citations)
- 2017 · Scalable Surface Reconstruction From Point Clouds With Extreme Scale and Density Diversity (CVPR 2017 · 36 citations)
- 2018 · PU-Net: Point Cloud Upsampling Network (CVPR 2018 · 689 citations)
- 2019 · Patch-Based Progressive 3D Point Set Upsampling (CVPR 2019 · 317 citations)
- 2020 · Deep Iterative Surface Normal Estimation (CVPR 2020 · 85 citations)
- 2021 · Point Cloud Upsampling via Disentangled Refinement (CVPR 2021 · 178 citations)
- 2022 · Neural Points: Point Cloud Representation With Neural Fields for Arbitrary Upsampling (CVPR 2022 · 91 citations)
- 2023 · Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent With Learned Distance Functions (CVPR 2023 · 94 citations)
- 2024 · A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · 32 citations)
- 2025 · 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · 18 citations)
- 2026 · PUFM: Efficient Point Cloud Upsampling via Flow Matching (AAAI 2026 · 4 citations)
Earliest papers here
- Sorting out typicality with the inverse moment matrix SOS polynomial (NeurIPS 2016 · 31 citations)
- Scalable Surface Reconstruction From Point Clouds With Extreme Scale and Density Diversity (CVPR 2017 · 36 citations)
- PU-Net: Point Cloud Upsampling Network (CVPR 2018 · 689 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.
PU-GAN 4× · CD (10^-3) (lower is better)
5 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 0.245 | Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent With Learned Distance Functions (CVPR 2023 · 94 citations) |
| 2024 | 0.131 | A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · 32 citations) |
| 2025 | 0.119 | Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · 4 citations) |
PU-GAN 4× · HD (10^-3) (lower is better)
5 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 3.08 | Neural Points: Point Cloud Representation With Neural Fields for Arbitrary Upsampling (CVPR 2022 · 91 citations) |
| 2023 | 2.369 | Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent With Learned Distance Functions (CVPR 2023 · 94 citations) |
| 2024 | 1.22 | A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · 32 citations) |
PU-GAN 4× · P2F (10^-3) (lower is better)
5 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 1.59 | Neural Points: Point Cloud Representation With Neural Fields for Arbitrary Upsampling (CVPR 2022 · 91 citations) |
| 2024 | 1.338 | Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · 26 citations) |
PU1K 4× · CD (10^-3) (lower is better)
5 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 0.404 | Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent With Learned Distance Functions (CVPR 2023 · 94 citations) |
| 2024 | 0.217 | A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · 32 citations) |
| 2025 | 0.155 | Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · 4 citations) |
PU-GAN 16× · HD (×10^-3) (lower is better)
4 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 6.52 | Neural Points: Point Cloud Representation With Neural Fields for Arbitrary Upsampling (CVPR 2022 · 91 citations) |
| 2023 | 2.352 | Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent With Learned Distance Functions (CVPR 2023 · 94 citations) |
| 2024 | 1.12 | A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · 32 citations) |
PU-GAN 16× · P2F (×10^-3) (lower is better)
4 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 1.49 | Neural Points: Point Cloud Representation With Neural Fields for Arbitrary Upsampling (CVPR 2022 · 91 citations) |
What has been tried
52 papers here report 560 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 3%, hurt 20%, 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 dual mapping supervision | PUGAN 4× · CD × 10−3 | 0.168 -> 0.131 | Helped | A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · 32 citations) |
| Use deformable rather than rigid RepKPoints | PU1K 4× · CD (×10^3) | 0.345 -> 0.327 | Helped | RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation (CVPR 2024 · 29 citations) |
| Replace inverse-distance interpolation with attention weighting | PU1K subset · CD (10^-3) | 0.759 -> 0.678 | Helped | Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · 26 citations) |
| Use coupled VelocityModules with a DistanceModule rather than a single VelocityModule without it | PUNet, 10K, 3% noise · CD | 3.73 -> 3.29 | Helped | StraightPCF: Straight Point Cloud Filtering (CVPR 2024 · 20 citations) |
| add differentiable rendering loss | PU-Net 50K, 2.5% noise · CD | 10.15 -> 9.28 | Helped | 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · 18 citations) |
| Remove RFA | PU1K sparse (512) · CD (10^-3) | 1.690 -> 1.722 | Hurt | SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · 18 citations) |
| Add CND-modified loss | PCPNet · CND (Ave.) | 8.41 -> 8.35 | Helped | CMG-Net: Robust Normal Estimation for Point Clouds via Chamfer Normal Distance and Multi-Scale Geometry (AAAI 2024 · 11 citations) |
| Use density-guided FPS instead of vanilla FPS | PU1K, 4× upsampling · CD (×10³) | 0.366 -> 0.352 | Helped | Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · 10 citations) |
| Use sparse column square-divided indexing (mode vii) rather than sparse column sum (mode v) | ModelNet40 · OA (%) | 93.40 -> 94.08 | Helped | SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · 6 citations) |
| add the adaptive classifier | PUNet 10K, 1%, 2%, and 3% noise · CD and P2M | CD: 1.973 -> 1.880; 2.794 -> 2.581; 3.580 -> 3.079; P2M: 0.230 -> 0.219; 0.555 -> 0.515; 1.108 -> 0.946 | Helped | You Should Learn to Stop Denoising on Point Clouds in Advance (AAAI 2025 · 5 citations) |
| Add normal-consistency loss Lnc | Thingi10K · CD (10−5) | 0.9643 -> 0.8551 | Helped | High-Fidelity Lightweight Mesh Reconstruction from Point Clouds (CVPR 2025 · 5 citations) |
| Set α:β to higher diversity weighting (1:1 to 10:1) | PartNet.v0.Merged · TMD | 4.36 -> 7.68 | Helped | Point Cloud Part Editing: Segmentation, Generation, Assembly, and Selection (AAAI 2024 · 5 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 |
|---|---|---|---|---|
| Grad-PU | PUGAN 16× | CD × 10−3 | 0.108 – 0.129 | 2 papers, e.g. CVPR 2024 and CVPR 2024 |
| Grad-PU | PU-GAN 4× | CD (10^-3) | 0.245 – 0.26 | 2 papers, e.g. CVPR 2024 and AAAI 2024 |
What reviewers will expect
Datasets and baselines named in the 18 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- PUNet · 3 of 18 papers
- Paris-rue-Madame · 3 of 18 papers
- ShapeNet · 3 of 18 papers
- PU-GAN · 2 of 18 papers
- PU1K · 2 of 18 papers
- PCPNet · 2 of 18 papers
- Thingi10K · 2 of 18 papers
Baselines
- PU-Net · 4 of 18 papers
- GPDNet · 3 of 18 papers
- MPU · 3 of 18 papers
- PU-GAN · 3 of 18 papers
- PU-GCN · 3 of 18 papers
- Grad-PU · 3 of 18 papers
- SHS-Net · 3 of 18 papers
- IterativePFN · 2 of 18 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 48 papers; from OpenAlex and the papers' affiliations).
- Nanyang Technological University · 4 papers
- Fudan University · 3 papers
- Nanjing University · 3 papers
- Australian National University · 3 papers
- Hikvision Research Institute · 2 papers
- Wayne State University · 2 papers
- Tsinghua University · 2 papers
- Deakin University · 2 papers
- Tencent · 2 papers
- University of British Columbia · 2 papers
Use it
Download BibTeX of 37 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
- Unsupervised shape correspondence (102 papers)
- Distance-function surface reconstruction (50 papers)
- Point cloud diffusion generation (44 papers)
Elsewhere in the topic
- Point cloud completion networks (50 papers)
- Structure-guided completion (22 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).