Research map / Image generation / Fast diffusion sampling
Diffusion theory and learned samplers
Provides variational and non-asymptotic views of diffusion and learns diffusions for sampling.
121 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Fast diffusion sampling, keyword cluster “score · distribution · denoising diffusion”
State History Tried Baselines 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.
40 papers in 2023–24 and 70 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 8.7% to 5.3%.
Open questions the data points to
- Papers report DSBM* on EMNIST → MNIST (FID) anywhere from 6 to 10.59: comparisons against it depend on whose number you take.
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
- 2020 · On the Ergodicity, Bias and Asymptotic Normality of Randomized Midpoint Sampling Method (NeurIPS 2020 · 7 citations)
- 2021 · Diffusion Schrödinger Bridge with Applications to Score-Based Generative Modeling (NeurIPS 2021 · 61 citations)
- 2022 · Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models (ICLR 2022 · 74 citations)
- 2023 · End-To-End Latent Variational Diffusion Models for Inverse Problems in High Energy Physics (NeurIPS 2023 · 19 citations)
- 2024 · On Diffusion Modeling for Anomaly Detection (ICLR 2024 · 4 citations)
- 2025 · A Periodic Bayesian Flow for Material Generation (ICLR 2025 · 2 citations)
- 2026 · Enhancing Diffusion-Based Sampling with Molecular Collective Variables (ICLR 2026 · 1 citation)
Earliest papers here
- On the Ergodicity, Bias and Asymptotic Normality of Randomized Midpoint Sampling Method (NeurIPS 2020 · 7 citations)
- Exact and Efficient Inference for Collective Flow Diffusion Model via Minimum Convex Cost Flow Algorithm (AAAI 2020 · 4 citations)
- Diffusion Schrödinger Bridge with Applications to Score-Based Generative Modeling (NeurIPS 2021 · 61 citations)
What has been tried
76 papers here report 667 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 7%, hurt 19%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use VicReg pre-training instead of no pre-training | CIFAR10, semi-supervised · AUC ROC | 51.53(14.81) -> 80.92(10.81) | Helped | On Diffusion Modeling for Anomaly Detection (ICLR 2024 · 4 citations) |
| Use time-scaled sampling instead of uniformly distributed time | time-scaled training objective experiment · test NLL | Helped | Generative Modeling on Manifolds Through Mixture of Riemannian Diffusion Processes (ICML 2024 · 3 citations) | |
| add EMA during online finetuning | AFHQ-64 Wild → Cat · FID | 32.665\std{0.647} -> 28.752\std{1.191} | Helped | Schrodinger Bridge Flow for Unpaired Data Translation (NeurIPS 2024 · 3 citations) |
| Increase Fourier basis size from 8 to 16 | Brownian motion · RMSE | 5.09 -> 6.66 | Hurt | Conditioning non-linear and infinite-dimensional diffusion processes (NeurIPS 2024 · 3 citations) |
| Replace Gaussian with non-Gaussian forward marginals | CIFAR-10 test · BPD | 2.49 -> 2.48 | Helped | Neural Flow Diffusion Models: Learnable Forward Process for Improved Diffusion Modelling (NeurIPS 2024 · 3 citations) |
| Use one expert instead of multiple experts | Avoiding · task entropy | Hurt | Variational Distillation of Diffusion Policies into Mixture of Experts (NeurIPS 2024 · 3 citations) | |
| add entropy conditioning instead of time conditioning | MP-20 stable structure prediction · Match rate (%) | 52.16 -> 64.35 | Helped | A Periodic Bayesian Flow for Material Generation (ICLR 2025 · 2 citations) |
| Use log-variance loss instead of KL loss | GMM (d=2) · ESS | 0.0051 -> 0.9093 | Helped | Improved sampling via learned diffusions (ICLR 2024 · 2 citations) |
| use Reciprocal projection instead of storing trajectory sample pairs | LJ-13 · E(·) W2 | 2.91 ±1.39 -> 2.40 ±1.25 | Helped | Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching (ICML 2025 · 1 citation) |
| Add RDKit warm-start to ASBS | Amortized conformer generation · recall coverage | Helped | Adjoint Schrödinger Bridge Sampler (NeurIPS 2025 · 1 citation) | |
| add relaxation | AL3 · Rama-KL | 8.535 ± 0.254 -> 1.209 ± 0.144 | Helped | Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities (NeurIPS 2025 · 1 citation) |
| Use log-variance divergence instead of KL divergence | Alanine Dipeptide · RMSD | Helped | Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers (ICLR 2025 · 1 citation) |
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 |
|---|---|---|---|---|
| DSBM* | EMNIST → MNIST | FID | 6 – 10.59 | 2 papers, e.g. NeurIPS 2024 and NeurIPS 2024 |
What reviewers will expect
Datasets and baselines named in the 45 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 · 8 of 45 papers
- MNIST · 6 of 45 papers
- FFHQ · 3 of 45 papers
- SPICE · 2 of 45 papers
- GEOM-DRUGS · 2 of 45 papers
- text8 · 2 of 45 papers
- AFHQv2 · 2 of 45 papers
- MW54 · 2 of 45 papers
Baselines
- SMC · 6 of 45 papers
- PIS · 4 of 45 papers
- DDS · 4 of 45 papers
- iDEM · 4 of 45 papers
- DPS · 4 of 45 papers
- D3PM · 3 of 45 papers
- RDKit ETKDG · 2 of 45 papers
- DDPM · 2 of 45 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 110 papers; from OpenAlex and the papers' affiliations).
- New York University · 7 papers
- Meta · 6 papers
- University of Oxford · 6 papers
- Georgia Institute of Technology · 5 papers
- University of California, Berkeley · 5 papers
- Korea Advanced Institute of Science and Technology · 5 papers
- Duke University · 5 papers
- Stanford University · 4 papers
- Université de Montréal · 4 papers
- Microsoft · 4 papers
Use it
Download BibTeX of 39 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 theory and design space (252 papers)
- Distributional control of diffusion (109 papers)
- Denoising diffusion probabilistic models (97 papers)
- Guidance for conditional diffusion (123 papers)
- Diffusion posterior sampling (72 papers)
- Diffusion models as perception tools (198 papers)
Elsewhere in the topic
- Training-free diffusion acceleration (257 papers)
- Masked diffusion language model decoding (147 papers)
- Fast diffusion ODE solvers (120 papers)
- Flow matching distillation (95 papers)
- Distillation for few-step generation (84 papers)
- Rectified flow straightening (70 papers)
- Discrete diffusion solvers (64 papers)
- Guidance and alignment for flow matching (34 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).