Research map / Image generation / Fast diffusion sampling
Discrete diffusion solvers
Develop fast solvers, correctors, and hybrid formulations for discrete-state diffusion.
64 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Fast diffusion sampling, keyword cluster “discrete · diffusion language · masked”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 1 tracked benchmark.
12 papers in 2023–24 and 50 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 2.6% to 3.8%.
Open questions the data points to
- CIFAR-10 (FID): best went from 4.77 (up to 2024) to 3.26 in 2025 — Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · 0 citations)
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 · Structured Denoising Diffusion Models in Discrete State-Spaces (NeurIPS 2021 · 82 citations)
- 2022 · Autoregressive Diffusion Models (ICLR 2022 · 1 citation)
- 2023 · Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning (ICLR 2023 · 80 citations)
- 2024 · Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · 15 citations)
- 2025 · Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · 6 citations)
- 2026 · Fine-Tuning Masked Diffusion for Provable Self-Correction (ICML 2026 · 0 citations)
Earliest papers here
- Structured Denoising Diffusion Models in Discrete State-Spaces (NeurIPS 2021 · 82 citations)
- Autoregressive Diffusion Models (ICLR 2022 · 1 citation)
- Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning (ICLR 2023 · 80 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)
4 papers, 2021–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 7.34 | Structured Denoising Diffusion Models in Discrete State-Spaces (NeurIPS 2021 · 82 citations) |
| 2023 | 4.77 | Blackout Diffusion: Generative Diffusion Models in Discrete-State Spaces (ICML 2023 · 1 citation) |
| 2025 | 3.26 | Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · 0 citations) |
What has been tried
40 papers here report 408 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 7%, hurt 20%, 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 |
|---|---|---|---|---|
| Use a cosine schedule instead of a linear schedule with a uniform sampling grid | ImageNet 64×64 · FID | 70 to 17 | Helped | Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · 15 citations) |
| Increase duplication size M | all domains · reward score | Mixed | Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · 6 citations) | |
| Replace score entropy with the concrete score matching ℓ² loss | Language modeling · perplexity | 10,000× higher perplexity | Hurt | Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution (ICML 2024 · 4 citations) |
| Use DDPP-LB rather than DDPP-IS | Amazon reviews · Gen PPL | 152.85 ± 1.64 -> 167.25 ± 27.33 | Hurt | Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction (ICLR 2025 · 2 citations) |
| Replace exact guidance with Taylor-approximated guidance (TAG) | QMugs-derived molecules, conditioned on ring count or LogP · property MAE | Mixed | Unlocking Guidance for Discrete State-Space Diffusion and Flow Models (ICLR 2025 · 2 citations) | |
| Add top-k token generation | Machine translation benchmarks · BLEU | approximately 1-2 points | Helped | Fast Sampling via Discrete Non-Markov Diffusion Models with Predetermined Transition Time (NeurIPS 2024 · 2 citations) |
| Add remasking ability and the proposed sampling improvements | OpenWebText · MAUVE | Helped | Remasking Discrete Diffusion Models with Inference-Time Scaling (NeurIPS 2025 · 1 citation) | |
| Add the relaxed objective during training | RADD, 8–64 NFEs · generative perplexity | 95.943 -> 89.830 (8 NFEs); 38.192 -> 36.106 (16 NFEs); 34.983 -> 33.234 (32 NFEs); 31.392 -> 29.115 (64 NFEs) | Helped | Learnable Sampler Distillation for Discrete Diffusion Models (NeurIPS 2025 · 1 citation) |
| Add logit temperature 0.6 | ImageNet 256×256 · FID | 29.12 -> 4.86 | Helped | Think while You Generate: Discrete Diffusion with Planned Denoising (ICLR 2025 · 1 citation) |
| Mix online and off-policy training instead of training online only | halfcheetah-medium-replay · average reward | 46.88±0.51 -> 48.11±0.56 | Helped | Amortizing intractable inference in diffusion models for vision, language, and control (NeurIPS 2024 · 1 citation) |
| include positional encodings | SketchGraphs test set · Bits/Sketch | 81.33 -> 83.03 | Hurt | SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation (ICML 2025 · 0 citations) |
| Add an informed corrector to HollowDiff | ImageNet 256×256 · FID | 10.24 -> 6.26 | Helped | Informed Correctors for Discrete Diffusion Models (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 30 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- OpenWebText · 14 of 30 papers
- CIFAR-10 · 10 of 30 papers
- ImageNet · 5 of 30 papers
- text8 · 4 of 30 papers
- QM9 · 4 of 30 papers
- LM1B · 3 of 30 papers
- WebText · 2 of 30 papers
- ZINC250k · 2 of 30 papers
Baselines
- MDLM · 11 of 30 papers
- SEDD · 7 of 30 papers
- D3PM · 6 of 30 papers
- Plaid · 5 of 30 papers
- MaskGIT · 5 of 30 papers
- GPT-2 · 4 of 30 papers
- UDLM · 4 of 30 papers
- MD4 · 3 of 30 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 62 papers; from OpenAlex and the papers' affiliations).
- NVIDIA · 4 papers
- Mila · 3 papers
- Stanford University · 3 papers
- Google DeepMind · 3 papers
- New York University · 3 papers
- Google · 2 papers
- University of Toronto · 2 papers
- Georgia Institute of Technology · 2 papers
- Los Alamos National Laboratory · 2 papers
- Hong Kong University of Science and Technology · 2 papers
Use it
Download BibTeX of 27 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
- Masked diffusion language model decoding (147 papers)
Elsewhere in the topic
- Training-free diffusion acceleration (257 papers)
- Diffusion theory and design space (252 papers)
- Diffusion models as perception tools (198 papers)
- Guidance for conditional diffusion (123 papers)
- Diffusion theory and learned samplers (121 papers)
- Fast diffusion ODE solvers (120 papers)
- Distributional control of diffusion (109 papers)
- Denoising diffusion probabilistic models (97 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).