Research map / Image generation / Fast diffusion sampling
Training-free diffusion acceleration
Speed diffusion inference through caching, pruning, early exit, and adaptive step selection.
257 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Fast diffusion sampling, keyword cluster “steps · distillation · time”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 3 tracked benchmarks.
51 papers in 2023–24 and 203 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 11.1% to 15.4%.
Open questions the data points to
- CIFAR-10 (FID): best went from 1.74 (up to 2024) to 1.54 in 2025 — Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · 0 citations)
- Papers report DDIM on CIFAR-10 (FID) anywhere from 4.16 to 13.65: comparisons against it depend on whose number you take.
What looks settled
- ImageNet 256×256 (FID): best up to 2024 1.56, best in 2025–26 1.99 from 5 papers, at most 0.2% better, if at all.
No longer reported by papers here: ImageNet 64×64 (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
- 2019 · Efficient Image Retrieval via Decoupling Diffusion into Online and Offline Processing (AAAI 2019 · 58 citations)
- 2022 · Progressive Distillation for Fast Sampling of Diffusion Models (ICLR 2022 · 186 citations)
- 2023 · On Distillation of Guided Diffusion Models (CVPR 2023 · 302 citations)
- 2024 · One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · 146 citations)
- 2025 · Efficient Diffusion as Low Light Enhancer (CVPR 2025 · 17 citations)
- 2026 · Forecast Then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers (AAAI 2026 · 3 citations)
Earliest papers here
- Efficient Image Retrieval via Decoupling Diffusion into Online and Offline Processing (AAAI 2019 · 58 citations)
- Progressive Distillation for Fast Sampling of Diffusion Models (ICLR 2022 · 186 citations)
- Tackling the Generative Learning Trilemma with Denoising Diffusion GANs (ICLR 2022 · 141 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)
15 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 2.57 | Progressive Distillation for Fast Sampling of Diffusion Models (ICLR 2022 · 186 citations) |
| 2024 | 1.74 | Directly Denoising Diffusion Models (ICML 2024 · 2 citations) |
| 2025 | 1.54 | Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · 0 citations) |
ImageNet 256×256 · FID (lower is better)
7 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 1.56 | PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher (NeurIPS 2024 · 1 citation) |
ImageNet 64×64 · FID (lower is better)
4 papers, 2023–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 4.7 | Consistency Models (ICML 2023 · 23 citations) |
| 2024 | 1.21 | PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher (NeurIPS 2024 · 1 citation) |
What has been tried
126 papers here report 1,448 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 6%, hurt 23%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add distribution matching objective | CIFAR-10 · FID | 3.82 -> 2.66 | Helped | One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · 146 citations) |
| Increase uniform caching interval from N=5 to N=10 | ImageNet 256 × 256 · FID | 3.59 -> 4.41 | Hurt | DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · 136 citations) |
| Increase the backbone feature scaling factor b | Image generation · Generation quality | Helped | FreeU: Free Lunch in Diffusion U-Net (CVPR 2024 · 135 citations) | |
| Apply minor training improvements, including hyperparameter tuning and disabling 32×32 self-attention | ImageNet-512 · FID | 8.00 to 7.24 | Helped | Analyzing and Improving the Training Dynamics of Diffusion Models (CVPR 2024 · 77 citations) |
| Remove warm-up steps | 10-step DPM-Solver, COCO Captions 2014 · LPIPS | 0.196 -> 0.404 | Hurt | DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models (CVPR 2024 · 44 citations) |
| Add scale-shift adjustment to block caching | COCO subset (faces removed), 20 steps · FID | 17.58 -> 15.95 | Helped | Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · 35 citations) |
| Remove Downsample Blocks 1 | Text-to-image generation · generation quality | Hurt | InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · 20 citations) | |
| Add reflectance-aware trajectory refinement | LOLv1, LOLv2-real, and LOLv2-synthetic · PSNR | Helped | Efficient Diffusion as Low Light Enhancer (CVPR 2025 · 17 citations) | |
| Add group size 128 to W4A8 RTN | ImageNet 256×256, 100 steps, cfg 1.5 · FID | 225.50 -> 13.77 | Helped | Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · 16 citations) |
| Add the adversarial loss | COCO2017, 2 NFEs · FID | Helped | Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025 · 15 citations) | |
| Add distillation PF-ODE | Super-resolution (DF2K) · FID | 20.307 -> 19.637 | Helped | CoDi: Conditional Diffusion Distillation for Higher-Fidelity and Faster Image Generation (CVPR 2024 · 15 citations) |
| Add spatially variant noise training and spatial noise embedding instead of using RAD reverse steps on an unmodified pretrained ADM | FFHQ Box · FID | 128.3 -> 22.1 | Helped | RAD: Region-Aware Diffusion Models for Image Inpainting (CVPR 2025 · 14 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 |
|---|---|---|---|---|
| DDIM | CIFAR-10 | FID | 4.16 – 13.65 | 2 papers, e.g. CVPR 2024 and CVPR 2024 |
What reviewers will expect
Datasets and baselines named in the 103 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet · 25 of 103 papers
- CIFAR-10 · 22 of 103 papers
- FFHQ · 10 of 103 papers
- MS-COCO · 10 of 103 papers
- ImageNet-1K · 7 of 103 papers
- LSUN-Church · 6 of 103 papers
- ImageNet 64x64 · 6 of 103 papers
- MS-COCO 2017 · 5 of 103 papers
Baselines
- DeepCache · 11 of 103 papers
- DDIM · 10 of 103 papers
- Q-Diffusion · 6 of 103 papers
- DiT-XL/2 · 6 of 103 papers
- SDXL-Lightning · 5 of 103 papers
- SDXL-Turbo · 5 of 103 papers
- DPM-Solver++ · 5 of 103 papers
- LCM · 4 of 103 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 254 papers; from OpenAlex and the papers' affiliations).
- Shanghai Jiao Tong University · 16 papers
- Tsinghua University · 11 papers
- Korea Advanced Institute of Science and Technology · 9 papers
- National University of Singapore · 9 papers
- Alibaba Group · 8 papers
- Zhejiang University · 8 papers
- ByteDance · 7 papers
- Google · 7 papers
- Nanyang Technological University · 7 papers
- Massachusetts Institute of Technology · 6 papers
Use it
Download BibTeX of 29 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
- Fast diffusion ODE solvers (120 papers)
- Distillation for few-step generation (84 papers)
- Accelerated diffusion policies (23 papers)
Elsewhere in the topic
- Diffusion theory and design space (252 papers)
- Diffusion models as perception tools (198 papers)
- Masked diffusion language model decoding (147 papers)
- Guidance for conditional diffusion (123 papers)
- Diffusion theory and learned samplers (121 papers)
- Distributional control of diffusion (109 papers)
- Denoising diffusion probabilistic models (97 papers)
- Flow matching distillation (95 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).