Research map / Image generation / Fast diffusion sampling
Fast diffusion ODE solvers
Reduce sampling steps with exponential integrators, parallel sampling, and optimized timesteps.
120 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Fast diffusion sampling, keyword cluster “steps · distillation · time”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 2 tracked benchmarks.
45 papers in 2023–24 and 72 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 9.8% to 5.4%.
Open questions the data points to
- CIFAR10 (FID): best went from 2.55 (up to 2024) to 2.14 in 2025 — Learning to Integrate Diffusion ODEs by Averaging the Derivatives (NeurIPS 2025 · 0 citations)
What looks settled
- LSUN-Bedroom (FID): best up to 2024 5.289, best in 2025–26 5.84 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
- 2022 · Dynamic Dual-Output Diffusion Models (CVPR 2022 · 24 citations)
- 2023 · Post-Training Quantization on Diffusion Models (CVPR 2023 · 138 citations)
- 2024 · Diffusion Models Without Attention (CVPR 2024 · 30 citations)
- 2025 · Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · 13 citations)
- 2026 · TR-DQ: Time-Rotation Diffusion Quantization (AAAI 2026 · 1 citation)
Earliest papers here
- Dynamic Dual-Output Diffusion Models (CVPR 2022 · 24 citations)
- Learning Fast Samplers for Diffusion Models by Differentiating Through Sample Quality (ICLR 2022 · 22 citations)
- Deep Equilibrium Approaches to Diffusion Models (NeurIPS 2022 · 7 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.
CIFAR10 · FID (lower is better)
12 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 2.55 | Fast Sampling of Diffusion Models with Exponential Integrator (ICLR 2023 · 57 citations) |
| 2025 | 2.14 | Learning to Integrate Diffusion ODEs by Averaging the Derivatives (NeurIPS 2025 · 0 citations) |
LSUN-Bedroom · FID (lower is better)
4 papers, 2024–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 5.289 | Denoising Diffusion Step-aware Models (ICLR 2024 · 3 citations) |
What has been tried
84 papers here report 887 measured changes (a component added, removed or swapped, compared with a baseline): helped 75%, no clear effect 4%, hurt 17%, 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 |
|---|---|---|---|---|
| use bidirectional SSMs instead of unidirectional SSMs | diffusion training loss · loss | Helped | Diffusion Models Without Attention (CVPR 2024 · 30 citations) | |
| use EDM instead of DPM-Solver++(3M) as AMED-Solver's teacher | CIFAR10 32×32, 5 NFE · FID | 12.34 -> 7.59 | Helped | Fast ODE-based Sampling for Diffusion Models in Around 5 Steps (CVPR 2024 · 28 citations) |
| Add temporal information aware reconstruction (TIAR) | LSUN-Bedrooms 256 × 256 · FID | 9.36 -> 4.84 | Helped | TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models (CVPR 2024 · 17 citations) |
| Add shuffling of denoising levels to iterative distillation | CIFAR10 · FID | Helped | DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture (CVPR 2025 · 12 citations) | |
| use self-attention rather than cross-attention maps for G-WPR | SD-XL image generation · image quality | Helped | Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models (CVPR 2024 · 11 citations) | |
| Use Pseudo-Huber distance instead of squared L2 distance | CIFAR-10 unconditional · FID | Helped | One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · 11 citations) | |
| Increase timesteps from 2 to 35 | ImageNet anisotropic deblurring · LPIPS | Helped | Deep Equilibrium Diffusion Restoration with Parallel Sampling (CVPR 2024 · 10 citations) | |
| Use first and last layer features instead of only first 2 layers as TPM inputs | COCO 2017 validation · Aesthetic v2 | 5.400 -> 5.445 | Helped | Schedule On the Fly: Diffusion Time Prediction for Faster and Better Image Generation (CVPR 2025 · 6 citations) |
| add discriminator loss | ImageNet 256p · FID | 4.32 -> 1.84 | Helped | Autoregressive Distillation of Diffusion Transformers (CVPR 2025 · 4 citations) |
| Add AIME auxiliary information mining and exploration | Foursquare · HR@10 (%) | 27.53 -> 39.88 | Helped | Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · 3 citations) |
| Vary the number of selected timesteps |S| used to approximate the objective change | CIFAR-10 · FID | Mixed | Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training (CVPR 2025 · 3 citations) | |
| Add input-signal denoising to SDSS | CIFAR-10 DDPM, 30–250 steps · FID | 7.93 -> 7.92; 3.55 -> 3.47; 3.09 -> 3.08 | Helped | Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models (AAAI 2024 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 51 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 · 16 of 51 papers
- ImageNet · 10 of 51 papers
- LSUN-Bedroom · 7 of 51 papers
- ImageNet 64x64 · 7 of 51 papers
- CelebA · 5 of 51 papers
- FFHQ · 5 of 51 papers
- MS-COCO · 5 of 51 papers
- ImageNet 256x256 · 3 of 51 papers
Baselines
- DDIM · 7 of 51 papers
- DPM-Solver · 6 of 51 papers
- UniPC · 5 of 51 papers
- DMD2 · 4 of 51 papers
- DPM-Solver-v3 · 4 of 51 papers
- Progressive Distillation · 4 of 51 papers
- iPNDM · 3 of 51 papers
- DPM-Solver-2 · 3 of 51 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 117 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 8 papers
- Zhejiang University · 6 papers
- NVIDIA · 5 papers
- Alibaba Group · 4 papers
- Nanjing University · 4 papers
- Shanghai Jiao Tong University · 4 papers
- Carnegie Mellon University · 4 papers
- Google · 4 papers
- Academy of Mathematics and Systems Science · 3 papers
- Huawei · 3 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
- Training-free diffusion acceleration (257 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).