Research map / 3D vision / 3D generation
Score distillation for text-to-3D
Distills 2D diffusion priors into 3D with multi-view consistency to produce text-conditioned assets.
106 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in 3D generation, keyword cluster “3d generation · text · consistent”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 3 tracked benchmarks, but the best result has barely moved since 2024.
50 papers in 2023–24 and 55 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 11.0% to 6.2%.
Open questions the data points to
- Papers report LGM on GSO (LPIPS) anywhere from 0.14 to 0.199: comparisons against it depend on whose number you take.
- Papers report LGM on GSO (PSNR) anywhere from 17.13 to 19.43: comparisons against it depend on whose number you take.
What looks settled
- GSO (PSNR): best up to 2024 26.07, best in 2025–26 23.02 from 2 papers, at most 0.2% better, if at all.
- GSO (SSIM): best up to 2024 0.924, best in 2025–26 0.884 from 2 papers, at most 0.2% better, if at all.
- GSO (LPIPS): best up to 2024 0.065, best in 2025–26 0.116 from 2 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
- 2022 · An Autoregressive Flow Model for 3D Molecular Geometry Generation from Scratch (ICLR 2022)
- 2023 · Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation (CVPR 2023 · 337 citations)
- 2024 · Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · 341 citations)
- 2025 · MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation (CVPR 2025 · 24 citations)
- 2026 · NaTex: Seamless Texture Generation as Latent Color Diffusion (CVPR 2026 · 1 citation)
Earliest papers here
- An Autoregressive Flow Model for 3D Molecular Geometry Generation from Scratch (ICLR 2022)
- Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation (CVPR 2023 · 337 citations)
- 3D Neural Field Generation Using Triplane Diffusion (CVPR 2023 · 164 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.
GSO · PSNR (higher is better)
4 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 26.07 | Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · 341 citations) |
GSO · SSIM (higher is better)
4 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.924 | Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · 341 citations) |
GSO · LPIPS (lower is better)
4 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 0.065 | Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · 341 citations) |
What has been tried
77 papers here report 714 measured changes (a component added, removed or swapped, compared with a baseline): helped 78%, no clear effect 2%, 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 |
|---|---|---|---|---|
| Remove cross-domain attention | Generated multi-view normal maps and color images · Cross-domain consistency | Hurt | Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · 341 citations) | |
| add noisy point growing and color perturbation | Sniper rifle prompt · visual detail | Helped | GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models (CVPR 2024 · 141 citations) | |
| Add the fine DMTet mesh stage after the coarse stage | Image-to-3D reconstruction · quantitative performance | Helped | Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors (ICLR 2024 · 74 citations) | |
| Add LAION 2D data during training | Multi-view image generation · FID | 40.38 -> 39.04 | Helped | MVDream: Multi-view Diffusion for 3D Generation (ICLR 2024 · 72 citations) |
| replace progressive pruning with a single pruning step | ShapeNet Chairs · grid IoU | 92.88% -> 89.68% | Hurt | XCube: Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies (CVPR 2024 · 69 citations) |
| Add Stable Diffusion guidance alongside Normal-Depth diffusion | Text-to-3D geometry ablation · Geometry CLIP Score | 24.1070 -> 25.8820 | Helped | RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D (CVPR 2024 · 66 citations) |
| Use normal-adapted and depth-adapted diffusion models instead of a text-to-image diffusion model for geometry generation | 3D human geometry generation · geometry quality | Helped | HumanNorm: Learning Normal Diffusion Model for High-quality and Realistic 3D Human Generation (CVPR 2024 · 46 citations) | |
| Generate 5 target views rather than 1 | In-domain diffusion samples · PSNR | 18.85 -> 21.66 | Helped | CAT3D: Create Anything in 3D with Multi-View Diffusion Models (NeurIPS 2024 · 39 citations) |
| Add semantic code sampling | Text-to-3D generation across viewpoints · semantic and geometric consistency | Helped | Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation (ICLR 2024 · 34 citations) | |
| Add a viewpoint-aware 3D prior | 3D geometry consistency · geometric consistency | Helped | DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 2024 · 25 citations) | |
| extend object self-attention with multi-instance attention (K=5) | 3D-Front · CD-S | 0.145 -> 0.080 | Helped | MIDI: Multi-Instance Diffusion for Single Image to 3D Scene Generation (CVPR 2025 · 24 citations) |
| use a video model instead of per-frame image inference | SyntheticObjects and SyntheticScenes · PSNR, SSIM, and LPIPS | 27.4 -> 28.3 PSNR; 25.4 -> 26.0 PSNR | Helped | Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models (CVPR 2025 · 23 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 |
|---|---|---|---|---|
| LGM | GSO | LPIPS | 0.14 – 0.199 | 2 papers, e.g. AAAI 2025 and CVPR 2025 |
| LGM | GSO | PSNR | 17.13 – 19.43 | 2 papers, e.g. CVPR 2025 and AAAI 2025 |
| LGM | GSO | SSIM | 0.808 – 0.854 | 2 papers, e.g. CVPR 2025 and AAAI 2025 |
What reviewers will expect
Datasets and baselines named in the 31 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 · 8 of 31 papers
- GSO · 5 of 31 papers
- 3D-FRONT · 4 of 31 papers
- G-Objaverse · 3 of 31 papers
- Objaverse LVIS · 2 of 31 papers
- ScanNet · 2 of 31 papers
- nuScenes · 2 of 31 papers
Baselines
- LGM · 5 of 31 papers
- DreamGaussian · 3 of 31 papers
- One-2-3-45 · 3 of 31 papers
- CRM · 3 of 31 papers
- InstantMesh · 3 of 31 papers
- 3DTopia · 3 of 31 papers
- Wonder3D · 2 of 31 papers
- TriplaneGaussian · 2 of 31 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 105 papers; from OpenAlex and the papers' affiliations).
- Nanyang Technological University · 9 papers
- Tsinghua University · 8 papers
- Zhejiang University · 7 papers
- Hong Kong University of Science and Technology · 6 papers
- ByteDance · 5 papers
- Shanghai AI Laboratory · 4 papers
- National University of Singapore · 4 papers
- University of Illinois, Urbana Champaign · 4 papers
- Tencent · 3 papers
- Johns Hopkins University · 3 papers
Use it
Download BibTeX of 24 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
- Diffusion-based novel view synthesis (156 papers)
- Feed-forward 3D asset generation and editing (116 papers)
- Text-to-3D scene generation (58 papers)
Elsewhere in the topic
- Learning 3D reconstruction from images (196 papers)
- Semantic scene completion (153 papers)
- Generative indoor scene synthesis (131 papers)
- Hand-object interaction generation (116 papers)
- Open-vocabulary 3D scene understanding (95 papers)
- 2D-to-3D pretrained representations (94 papers)
- Monocular 3D object reconstruction (94 papers)
- Scene-aware human motion synthesis (82 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).