Research map / Probabilistic and scientific modeling / Variational and generative sampling
Score-based generative modeling analysis
Estimates data gradients and analyzes bias, generalization and evaluation of generative models.
128 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Variational and generative sampling, keyword cluster “diffusion · score · matching”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports None of the tracked benchmarks has a report from 2025–26.
30 papers in 2023–24 and 38 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 7.2% to 7.6%.
No longer reported by papers here: CIFAR-10 (FID, 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 · A Theory of Generative ConvNet (ICML 2016 · 168 citations)
- 2017 · Compressed Sensing using Generative Models (ICML 2017 · 281 citations)
- 2018 · Assessing Generative Models via Precision and Recall (NeurIPS 2018 · 168 citations)
- 2019 · Generative Modeling by Estimating Gradients of the Data Distribution (NeurIPS 2019 · 1,076 citations)
- 2020 · Effectively Unbiased FID and Inception Score and Where to Find Them (CVPR 2020 · 227 citations)
- 2021 · Robust Compressed Sensing MRI with Deep Generative Priors (NeurIPS 2021 · 45 citations)
- 2022 · Wavelet Score-Based Generative Modeling (NeurIPS 2022 · 10 citations)
- 2023 · A Likelihood Approach to Nonparametric Estimation of a Singular Distribution Using Deep Generative Models (ICML 2023 · 13 citations)
- 2024 · Estimating the Hallucination Rate of Generative AI (NeurIPS 2024 · 9 citations)
- 2025 · Unveiling Differences in Generative Models: A Scalable Differential Clustering Approach (CVPR 2025 · 2 citations)
- 2026 · Temper-Then-Tilt: Principled Unlearning for Generative Models through Tempering and Classifier Guidance (ICML 2026 · 0 citations)
Earliest papers here
- A Theory of Generative ConvNet (ICML 2016 · 168 citations)
- Conditional Generative Moment-Matching Networks (NeurIPS 2016 · 35 citations)
- Compressed Sensing using Generative Models (ICML 2017 · 281 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.
CIFAR-10 · FID (lower is better)
7 papers, 2019–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 17.29 | Bias Correction of Learned Generative Models using Likelihood-Free Importance Weighting (NeurIPS 2019 · 65 citations) |
| 2021 | 2.1 | Score-based Generative Modeling in Latent Space (NeurIPS 2021 · 12 citations) |
What has been tried
100 papers here report 861 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 9%, hurt 21%, mixed 7%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| sample noise using real-data frequency statistics instead of standard Gaussian noise | CelebA and MNIST generation · generation quality | Helped | Idempotent Generative Network (ICLR 2024 · 4 citations) | |
| Add the scaling factor γ | 1-bit CS on MNIST, CIFAR-10, and CelebA with ill-conditioned sensing matrices · PSNR | Helped | QCS-SGM+: Improved Quantized Compressed Sensing with Score-Based Generative Models (AAAI 2024 · 3 citations) | |
| Use n_R=50 or n_R=200 instead of n_R=100 | CIFAR-10 generative models · χ²_PQM | No clear effect | PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation (ICLR 2025 · 2 citations) | |
| Add fast sampling (end-signal switching) to SSM | Linear, Quadratic, Log-Log Linear, Log-Log Cubic, Sinusoidal · inference time | Helped | Supervised Score-Based Modeling by Gradient Boosting (AAAI 2025 · 1 citation) | |
| Add training across the entire simplex rather than only on simplex edges | MNIST digits 0–5 · image translation quality | Helped | Multimarginal Generative Modeling with Stochastic Interpolants (ICLR 2024 · 1 citation) | |
| apply the truncation trick | FFHQ · KEN | Hurt | An Interpretable Evaluation of Entropy-based Novelty of Generative Models (ICML 2024 · 1 citation) | |
| Add Gaussian noise to MNIST training data | MNIST test images · W1 | Sparse: 1.9555 ± 0.7182 at 0 to 4.0880 ± 0.7518 at 0.5; fully connected: 1.8859 ± 0.7355 at 0 to 4.0995 ± 0.7633 at 0.5 | Hurt | A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · 0 citations) |
| Add GW-EP regularization | MNIST → USPS · accuracy | 88.1 -> 91.4 | Helped | Expressive Score-Based Priors for Distribution Matching with Geometry-Preserving Regularization (ICML 2025 · 0 citations) |
| Use two-stage rather than joint training | MNIST-Binary · epochs to convergence | 80 epochs -> 20 epochs | Helped | Generative Marginalization Models (ICML 2024 · 0 citations) |
| Add self-normalization to the baseline | Reddit TL;DR · Train RM win-rate | 61.4 -> 95.2 | Helped | BRAIn: Bayesian Reward-conditioned Amortized Inference for natural language generation from feedback (ICML 2024 · 0 citations) |
| increase diffusion steps from T=3 to T=6 | CIFAR-10 · FID | 8.93 -> 8.13 | Helped | Learning Latent Space Hierarchical EBM Diffusion Models (ICML 2024 · 0 citations) |
| Add variance reduction to first-order score estimation | 100-dimensional correlated Gaussian, MCAR, α=0.3, σ=0.01 · MSE | 0.44 ± 0.01 -> 0.15 ± 0.01 | Helped | MissScore: High-Order Score Estimation in the Presence of Missing Data (ICML 2025) |
What reviewers will expect
Datasets and baselines named in the 23 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 5 of 23 papers
- MNIST · 4 of 23 papers
- ImageNet · 3 of 23 papers
- CIFAR-100 · 2 of 23 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).
- New York University · 4 papers
- Google DeepMind · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Renmin University of China · 3 papers
- Stanford University · 3 papers
- Columbia University · 3 papers
- Huawei · 2 papers
- The University of Tokyo · 2 papers
- Nanyang Technological University · 2 papers
- Tsinghua University · 2 papers
Use it
Download BibTeX of 40 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
- Posterior sampling with diffusion guidance (137 papers)
- Flow matching and flow sampling (144 papers)
Elsewhere in the topic
- Likelihood-free and graphical model estimation (161 papers)
- Structured latent VAEs (125 papers)
- Exact inference for probabilistic programs (103 papers)
- Disentangled and discrete latent models (92 papers)
- Latent variable marginalization and estimators (91 papers)
- Tractable probabilistic models (90 papers)
- Variational inference theory and gradients (89 papers)
- Hamiltonian Monte Carlo (89 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).