Research map / Point clouds / Shape and surface completion
Point cloud diffusion generation
Generate 3D shapes with diffusion models over point clouds and latent spaces.
44 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 Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
11 papers in 2023–24 and 24 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 12.4% to 21.2%.
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 · Tweet Timeline Generation with Determinantal Point Processes (AAAI 2016 · 21 citations)
- 2018 · Towards a Neural Conversation Model With Diversity Net Using Determinantal Point Processes (AAAI 2018 · 29 citations)
- 2020 · SoftFlow: Probabilistic Framework for Normalizing Flow on Manifolds (NeurIPS 2020 · 51 citations)
- 2021 · Diffusion Probabilistic Models for 3D Point Cloud Generation (CVPR 2021 · 620 citations)
- 2022 · LION: Latent Point Diffusion Models for 3D Shape Generation (NeurIPS 2022 · 190 citations)
- 2023 · Fast Point Cloud Generation With Straight Flows (CVPR 2023 · 33 citations)
- 2024 · TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · 14 citations)
- 2025 · EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from Point Clouds (CVPR 2025 · 12 citations)
- 2026 · RaLD: Generating High-Resolution 3D Radar Point Clouds with Latent Diffusion (AAAI 2026 · 2 citations)
Earliest papers here
- Tweet Timeline Generation with Determinantal Point Processes (AAAI 2016 · 21 citations)
- Towards a Neural Conversation Model With Diversity Net Using Determinantal Point Processes (AAAI 2018 · 29 citations)
- SoftFlow: Probabilistic Framework for Normalizing Flow on Manifolds (NeurIPS 2020 · 51 citations)
What has been tried
28 papers here report 310 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 2%, hurt 29%, 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 |
|---|---|---|---|---|
| Replace timestep-dependent branch weights with fixed equal weights | ShapeNetv2, three-category mean · CD | 59.43 -> 64.45 | Hurt | TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · 14 citations) |
| Use endpoint parameterization instead of midpoint-offset | Entry-level dataset · EF1 | 0.88 -> 0.19 | Hurt | EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from Point Clouds (CVPR 2025 · 12 citations) |
| use block size 4 and offset size 32 instead of block size 8 and offset size 16 | Point-cloud conditional generation · Hausdorff distance | 0.166 -> 0.209 | Hurt | Scaling Mesh Generation via Compressive Tokenization (CVPR 2025 · 9 citations) |
| Use hierarchical coordinate-prediction heads instead of simultaneous x-y-z prediction | Objaverse evaluation set · CD | 0.0114 -> 0.0070 | Helped | TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing (CVPR 2025 · 9 citations) |
| Replace random feature initialization with zero initialization | Neural point feature space · Cosine sim. | 0.0306 -> 0.7695 | Helped | Neural Point Cloud Diffusion for Disentangled 3D Shape and Appearance Generation (CVPR 2024 · 7 citations) |
| train at resolution 64 instead of the default resolution 1024 | CO3D-v2, test resolution 16k · CD@1k | 0.179 -> 0.178 | Helped | PointInfinity: Resolution-Invariant Point Diffusion Models (CVPR 2024 · 5 citations) |
| increase model size from DiM-3D-S to DiM-3D-XL | ShapeNet Chair · 1-NNA CD | 57.27 -> 45.78 | Helped | Scaling Diffusion Mamba with Bidirectional SSMs for Efficient 3D Shape Generation (AAAI 2025 · 3 citations) |
| Add unlabeled training data to SeaLion trained with 10% labeled data | ShapeNet car · 1-NNA (p-CD) (%) | 87.34 -> 83.23 | Helped | SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation (CVPR 2025 · 3 citations) |
| Increase quantization resolution from 128^3 to 512^3 | Mesh generation quality · Surface quality | Helped | EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation (ICLR 2025 · 3 citations) | |
| increase octree depth from 10 to 11 | ShapeNet bench · FID | 26.06 -> 23.21 | Helped | TexOct: Generating Textures of 3D Models with Octree-based Diffusion (CVPR 2024 · 3 citations) |
| Add farthest-sampled attention | ABO test (100 shapes) · FID | 83.16 -> 77.14 | Helped | UV-free Texture Generation with Denoising and Geodesic Heat Diffusion (NeurIPS 2024 · 2 citations) |
| Add SLDM and LPGDM | MVUB attacker → 8iv2 victim, PCGCv2 · p2plane-PSNR (M2) | Helped | VVRec: Reconstruction Attacks on DL-based Volumetric Video Upstreaming via Latent Diffusion Model with Gamma Distribution (AAAI 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 13 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Objaverse · 4 of 13 papers
- ShapeNet · 3 of 13 papers
- Objaverse-XL · 2 of 13 papers
- GSO · 2 of 13 papers
- G-Objaverse · 2 of 13 papers
Baselines
- MeshAnything · 3 of 13 papers
- MeshAnythingV2 · 3 of 13 papers
- DiT-3D · 2 of 13 papers
- PVD · 2 of 13 papers
- LION · 2 of 13 papers
- Point-E · 2 of 13 papers
- EdgeRunner · 2 of 13 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 35 papers; from OpenAlex and the papers' affiliations).
- The Chinese University of Hong Kong · 2 papers
- Meta · 2 papers
- Tencent · 2 papers
Use it
Download BibTeX of 31 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)
- Point cloud upsampling and normals (64 papers)
- Distance-function surface reconstruction (50 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).