Research map / Image generation / Fast diffusion sampling
Diffusion theory and design space
Analyze convergence of score-based diffusion and optimize noise schedules and training design.
252 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Fast diffusion sampling, keyword cluster “score · distribution · denoising diffusion”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 2 tracked benchmarks.
66 papers in 2023–24 and 174 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 14.3% to 13.2%.
Open questions the data points to
- CelebA (FID): best went from 11.6 (up to 2024) to 2.97 in 2025 — CPSample: Classifier Protected Sampling for Guarding Training Data During Diffusion (ICLR 2025 · 0 citations)
- Papers report DDPM on CIFAR-10 (FID) anywhere from 1.97 to 4.04: comparisons against it depend on whose number you take.
- Papers report DDIM on CelebA (FID) anywhere from 1.27 to 6.53: comparisons against it depend on whose number you take.
What looks settled
- CIFAR-10 (FID): best up to 2024 1.77, best in 2025–26 2.81 from 5 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
- 2021 · Variational Diffusion Models (NeurIPS 2021 · 276 citations)
- 2022 · Elucidating the Design Space of Diffusion-Based Generative Models (NeurIPS 2022 · 301 citations)
- 2023 · Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC (ICML 2023 · 15 citations)
- 2024 · Plug-In Diffusion Model for Sequential Recommendation (AAAI 2024 · 63 citations)
- 2025 · Erasing Undesirable Influence in Diffusion Models (CVPR 2025 · 8 citations)
- 2026 · SIDE: Surrogate Conditional Data Extraction from Diffusion Models (AAAI 2026 · 1 citation)
Earliest papers here
- Variational Diffusion Models (NeurIPS 2021 · 276 citations)
- Argmax Flows and Multinomial Diffusion: Learning Categorical Distributions (NeurIPS 2021 · 36 citations)
- Learning Energy-Based Models by Diffusion Recovery Likelihood (ICLR 2021 · 15 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)
14 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 7.41 | Variational Diffusion Models (NeurIPS 2021 · 276 citations) |
| 2022 | 2.37 | Maximum Likelihood Training for Score-based Diffusion ODEs by High Order Denoising Score Matching (ICML 2022 · 11 citations) |
| 2023 | 1.77 | Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models (ICML 2023 · 13 citations) |
CelebA · FID (lower is better)
5 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 11.6 | Constrained Diffusion Models via Dual Training (NeurIPS 2024 · 1 citation) |
| 2025 | 2.97 | CPSample: Classifier Protected Sampling for Guarding Training Data During Diffusion (ICLR 2025 · 0 citations) |
What has been tried
158 papers here report 1,535 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 5%, hurt 21%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add HBR | Toy and Game · N@10 | Toy: 0.2261 -> 0.2311; Game: 0.3905 -> 0.3963 | Helped | Plug-In Diffusion Model for Sequential Recommendation (AAAI 2024 · 63 citations) |
| Replace the forgetting target with uniform noise instead of random-label predictions | CIFAR-10 (forget airplane) · FID | 8.66 -> 7.61 | Helped | Erasing Undesirable Influence in Diffusion Models (CVPR 2025 · 8 citations) |
| Use mean estimation instead of sample (DDIM) estimation | Darcy flow · Residual error | Helped | Physics-Informed Diffusion Models (ICLR 2025 · 8 citations) | |
| Use rotation-matrix inputs instead of axis-angle inputs | Partially trained sampled distributions · distribution continuity | Helped | Unified Framework for Diffusion Generative Models in SO(3): Applications in Computer Vision and Astrophysics (AAAI 2024 · 6 citations) | |
| increase particle count from 4 to 8 | GenEval and human preference alignment · score | Helped | A General Framework for Inference-time Scaling and Steering of Diffusion Models (ICML 2025 · 5 citations) | |
| Add higher-order bias correction and estimates w_t to the probability flow ODE sampler | Forward-process q1 vs generated p1 · TV distance convergence scaling in accuracy ε | \widetilde{O}(1/\varepsilon) -> \widetilde{O}(1/\sqrt{\varepsilon}) | Helped | Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models (ICLR 2024 · 5 citations) |
| Add bridge denoising to Gaussian prior actions | Indoor adaptation task · success rate | 0% -> 64% | Helped | Prior Does Matter: Visual Navigation via Denoising Diffusion Bridge Models (CVPR 2025 · 4 citations) |
| add observed-value conditioning to the forward process | METR-LA Block-missing · MAE | 2.05±0.01 -> 1.96±0.01 | Helped | RDPI: A Refine Diffusion Probability Generation Method for Spatiotemporal Data Imputation (AAAI 2025 · 3 citations) |
| Use classifier-free guidance with ω=0.6 instead of ω=0.0 | Seven datasets (average) · Improvement over XGBoost | 3.76 -> 5.60 | Helped | TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · 3 citations) |
| use latent-space distances instead of parameter-space distances | Rastrigin 32D · average fitness | 0.02 -> 0.19 | Helped | Diffusion Models are Evolutionary Algorithms (ICLR 2025 · 3 citations) |
| Set τ₂ to 0.5 instead of 0.9 | ImageNet class-conditional generation, CFG 5 · FID | 9.47 -> 8.21 | Helped | CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024 · 3 citations) |
| Increase the number of modulation-layer/timestep sets up to M=10 | Image generation · generation quality | Helped | Optical Diffusion Models for Image Generation (NeurIPS 2024 · 3 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 |
|---|---|---|---|---|
| DDPM | CIFAR-10 | FID | 1.97 – 4.04 | 3 papers, e.g. ICML 2025 and ICML 2025 |
| DDIM | CelebA | FID | 1.27 – 6.53 | 2 papers, e.g. ICLR 2025 and ICML 2025 |
| DDIM | CIFAR-10 | FID | 3.17 – 4.16 | 2 papers, e.g. ICLR 2025 and ICML 2025 |
What reviewers will expect
Datasets and baselines named in the 116 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 · 35 of 116 papers
- MNIST · 18 of 116 papers
- CelebA · 12 of 116 papers
- ImageNet · 12 of 116 papers
- FFHQ · 9 of 116 papers
- LSUN Bedroom · 4 of 116 papers
- COCO · 4 of 116 papers
- CelebA-HQ · 4 of 116 papers
Baselines
- DPS · 7 of 116 papers
- EDM · 6 of 116 papers
- DDIM · 6 of 116 papers
- DDPM · 4 of 116 papers
- DAS · 4 of 116 papers
- D-TRAK · 4 of 116 papers
- SDEdit · 3 of 116 papers
- DiT · 3 of 116 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 240 papers; from OpenAlex and the papers' affiliations).
- Massachusetts Institute of Technology · 10 papers
- Google · 8 papers
- Korea Advanced Institute of Science and Technology · 6 papers
- University of Oxford · 6 papers
- Carnegie Mellon University · 5 papers
- Shanghai Jiao Tong University · 5 papers
- Stanford University · 5 papers
- Imperial College London · 5 papers
- University of Texas, Austin · 5 papers
- Cornell University · 4 papers
Use it
Download BibTeX of 38 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
- Distributional control of diffusion (109 papers)
- Denoising diffusion probabilistic models (97 papers)
- Guidance for conditional diffusion (123 papers)
- Diffusion posterior sampling (72 papers)
- Diffusion theory and learned samplers (121 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).