Research map / Image generation / Autoregressive and latent generation
Equivariant diffusion for 3D molecules
Generate 3D molecular structures with equivariant diffusion in geometric or latent space.
71 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Autoregressive and latent generation, keyword cluster “protein · molecular · molecule”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 1 tracked benchmark, but the best result has barely moved since 2024.
25 papers in 2023–24 and 43 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 5.3% to 4.8%.
What looks settled
- QM9 (validity (%)): best up to 2024 98.6, best in 2025–26 97.43 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
- 2019 · Generating Chinese Ci with Designated Metrical Structure (AAAI 2019)
- 2022 · Equivariant Diffusion for Molecule Generation in 3D (ICML 2022 · 118 citations)
- 2023 · MDM: Molecular Diffusion Model for 3D Molecule Generation (AAAI 2023 · 81 citations)
- 2024 · Scalable Diffusion for Materials Generation (ICLR 2024 · 25 citations)
- 2025 · DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry (CVPR 2025 · 10 citations)
- 2026 · OXtal: An All-Atom Diffusion Model for Organic Crystal Structure Prediction (ICLR 2026 · 2 citations)
Earliest papers here
- Generating Chinese Ci with Designated Metrical Structure (AAAI 2019)
- Equivariant Diffusion for Molecule Generation in 3D (ICML 2022 · 118 citations)
- Torsional Diffusion for Molecular Conformer Generation (NeurIPS 2022 · 82 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.
QM9 · validity (%) (higher is better)
4 papers, 2023–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 98.6 | MDM: Molecular Diffusion Model for 3D Molecule Generation (AAAI 2023 · 81 citations) |
What has been tried
46 papers here report 499 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 5%, hurt 19%, 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 |
|---|---|---|---|---|
| Increase UniMat feature dimension | MP20 · Validity and coverage | Helped | Scalable Diffusion for Materials Generation (ICLR 2024 · 25 citations) | |
| replace DiffCSP's lattice inner-product representation with coefficient vector k | MP-20 · MR (%) | 51.49 -> 50.76 | Hurt | Space Group Constrained Crystal Generation (ICLR 2024 · 15 citations) |
| Add triplet-wise geometric information through the pair-triplet attention track | GEOM-QM9 · Mol Stability (%) | 85.9±0.2 -> 87.7±0.2 | Helped | Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · 12 citations) |
| Replace B-spline representations with point-cloud-based latent encoding | DeepCAD · Valid (%) | 79.80 -> 71.77 | Hurt | DTGBrepGen: A Novel B-rep Generative Model through Decoupling Topology and Geometry (CVPR 2025 · 10 citations) |
| connect all generated graph components instead of keeping only the largest connected component | multi-conditional molecular generation · condition controllability | Helped | Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · 9 citations) | |
| train jointly on QM9 and MP20 instead of MP20 only | MP20 · overall validity (%) | 90.13 -> 91.92 | Helped | All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · 7 citations) |
| Use asymmetric unit representation instead of a full conventional unit cell | MP-20 · training epoch time | 52 sec. -> 27 sec. | Helped | SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models (ICLR 2025 · 7 citations) |
| Use time-dependent rather than uniform loss weighting | GEOM-Drugs · Molecular stability | 87.59 -> 91.60 | Helped | Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · 7 citations) |
| replace the uniform prior with the empirical marginal prior | WBM · FWD | 2.29±0.15 -> 1.65±0.07 | Helped | WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · 3 citations) |
| decode the lattice from the noisy cell instead of a cubic cell | Materials Project test set · validity | 99.17\% -> 15.23\% | Hurt | Vector Field Oriented Diffusion Model for Crystal Material Generation (AAAI 2024 · 3 citations) |
| Add retrieval-augmented generation | Three crystal families · Valid formula rate | 0.97 -> 1.00 | Helped | Generative Hierarchical Materials Search (NeurIPS 2024 · 3 citations) |
| Add discrete latent space (DL) | PED00055, PED00090, PED00151, PED00218 · RMSD | 1.743 -> 1.736; 1.810 -> 1.806; 1.557 -> 1.549; 1.620 -> 1.595 | Helped | Exploit Your Latents: Coarse-Grained Protein Backmapping with Latent Diffusion Models (AAAI 2025 · 2 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 |
|---|---|---|---|---|
| EDM | GEOM_DRUG | QED | 0.558 – 0.608 | 2 papers, e.g. ICML 2023 and ICML 2023 |
What reviewers will expect
Datasets and baselines named in the 28 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GEOM-DRUGS · 7 of 28 papers
- MP-20 · 6 of 28 papers
- QM9 · 5 of 28 papers
- MPTS-52 · 4 of 28 papers
- Carbon-24 · 4 of 28 papers
- GEOM-QM9 · 3 of 28 papers
- Perov-5 · 3 of 28 papers
- Materials Project · 2 of 28 papers
Baselines
- CDVAE · 8 of 28 papers
- DiffCSP · 7 of 28 papers
- GeoLDM · 6 of 28 papers
- EDM · 5 of 28 papers
- FlowMM · 5 of 28 papers
- GeoDiff · 4 of 28 papers
- SemlaFlow · 4 of 28 papers
- EQGAT-diff · 4 of 28 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 68 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 7 papers
- Peking University · 6 papers
- Stanford University · 5 papers
- Massachusetts Institute of Technology · 4 papers
- Renmin University of China · 4 papers
- University of California, Berkeley · 3 papers
- Google · 3 papers
- Sun Yat-sen University · 3 papers
- City University of Hong Kong · 2 papers
- The Chinese University of Hong Kong · 2 papers
Use it
Download BibTeX of 30 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 for protein and antibody design (48 papers)
Elsewhere in the topic
- Classical generative image synthesis (147 papers)
- High-resolution diffusion synthesis (140 papers)
- Retrieval and matching with generative features (140 papers)
- Generative latent space tooling (129 papers)
- Generalizable fake image detection (119 papers)
- Latent diffusion with learned tokenizers (114 papers)
- Synthetic training data from generators (111 papers)
- Compositional scene generation (105 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).