Research map / Image generation / Fast diffusion sampling
Accelerated diffusion policies
Speed robotic and driving diffusion policies with distillation and pruning.
23 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Fast diffusion sampling, keyword cluster “steps · distillation · time”
State History Tried 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.
5 papers in 2023–24 and 18 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 1.1% to 1.4%.
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
- 2023 · Leapfrog Diffusion Model for Stochastic Trajectory Prediction (CVPR 2023 · 209 citations)
- 2024 · SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · 13 citations)
- 2025 · DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · 87 citations)
- 2026 · MP1: MeanFlow Tames Policy Learning in 1-step for Robotic Manipulation (AAAI 2026 · 2 citations)
Earliest papers here
- Leapfrog Diffusion Model for Stochastic Trajectory Prediction (CVPR 2023 · 209 citations)
- BCDiff: Bidirectional Consistent Diffusion for Instantaneous Trajectory Prediction (NeurIPS 2023 · 10 citations)
- SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · 13 citations)
What has been tried
12 papers here report 147 measured changes (a component added, removed or swapped, compared with a baseline): helped 73%, no clear effect 5%, hurt 18%, 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 |
|---|---|---|---|---|
| remove spatial cross-attention | NAVSIM navtest · PDMS | 88.1 -> 55.1 | Hurt | DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · 87 citations) |
| increase temporal resolution by decreasing temporal patch size | WOMD scene generation · realism | Helped | SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · 13 citations) | |
| use only the last planning level instead of all PRP levels | Hopper-me · normalized score | 111.6±0.2 -> 27.8±10.2 | Hurt | DiffuserLite: Towards Real-time Diffusion Planning (NeurIPS 2024 · 11 citations) |
| Add cascaded body-hand estimation rather than estimating whole-body keypoints jointly | ColossusEgo · MPJPE Body | 55.09 -> 53.83 | Helped | REWIND: Real-Time Egocentric Whole-Body Motion Diffusion with Exemplar-Based Identity Conditioning (CVPR 2025 · 3 citations) |
| add frequency consistency with adaptive frequency loss | Robomimic Transport · success rate | 0.84 -> 0.92 | Helped | FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · 2 citations) |
| Add Adaptive Chunking | Six simulation benchmarks · success rate | Helped | Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · 1 citation) | |
| use listwise preference loss instead of removing it | MovieLens-1M (len=5) · SeqHR@5 | 0.0000 -> 0.1218 | Helped | Listwise Preference Diffusion Optimization for User Behavior Trajectories Prediction (NeurIPS 2025 · 1 citation) |
| Use OneDP-S instead of OneDP-D | Robomimic and PushT simulation tasks · success rate or target-area coverage | Helped | One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · 0 citations) | |
| Reduce noise scale to gamma 0 | Adroit Hand tasks · normalized success rate | 0.68 -> 0.99 | Helped | Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · 0 citations) |
| replace diffusion purification with Gaussian filter | Ant-v2 D1-L1 · average cumulative reward | 242 vs 261 | Hurt | Imitation Learning from Purified Demonstrations (ICML 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 8 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Robomimic · 2 of 8 papers
- PushT · 2 of 8 papers
Baselines
- Consistency Policy · 2 of 8 papers
- DDPM · 2 of 8 papers
- DDIM · 2 of 8 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 23 papers; from OpenAlex and the papers' affiliations).
- Google · 2 papers
- University of California, San Diego · 2 papers
Use it
Download BibTeX of 22 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)
- Fast diffusion ODE solvers (120 papers)
- Distillation for few-step generation (84 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).