Research map / Image generation / Autoregressive and latent generation
Synthetic training data from generators
Use generated images for classification training and evaluate generative models.
111 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Autoregressive and latent generation, keyword cluster “captioning · assessment · synthetic”
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.
44 papers in 2023–24 and 35 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 9.3% to 3.9%.
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
- 2016 · Synthesizing the preferred inputs for neurons in neural networks via deep generator networks (NeurIPS 2016 · 249 citations)
- 2017 · Deliberation Networks: Sequence Generation Beyond One-Pass Decoding (NeurIPS 2017 · 192 citations)
- 2018 · Compatibility Family Learning for Item Recommendation and Generation (AAAI 2018 · 37 citations)
- 2019 · Improved Precision and Recall Metric for Assessing Generative Models (NeurIPS 2019 · 296 citations)
- 2020 · Reliable Fidelity and Diversity Metrics for Generative Models (ICML 2020 · 54 citations)
- 2021 · Auto-Encoding Knowledge Graph for Unsupervised Medical Report Generation (NeurIPS 2021 · 64 citations)
- 2022 · Kubric: A Scalable Dataset Generator (CVPR 2022 · 208 citations)
- 2023 · InCoder: A Generative Model for Code Infilling and Synthesis (ICLR 2023 · 133 citations)
- 2024 · Learning to Rank in Generative Retrieval (AAAI 2024 · 21 citations)
- 2025 · Unseen Visual Anomaly Generation (CVPR 2025 · 21 citations)
- 2026 · Is Generation Required for Data-Efficient Perception? (ICML 2026 · 0 citations)
Earliest papers here
- Synthesizing the preferred inputs for neurons in neural networks via deep generator networks (NeurIPS 2016 · 249 citations)
- One-Shot Generalization in Deep Generative Models (ICML 2016 · 74 citations)
- Deliberation Networks: Sequence Generation Beyond One-Pass Decoding (NeurIPS 2017 · 192 citations)
What has been tried
91 papers here report 979 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 5%, hurt 17%, 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 test-time normal sample conditioning | Anomaly generation · generation quality | Helped | Unseen Visual Anomaly Generation (CVPR 2025 · 21 citations) | |
| Remove both margin-based rank losses | Natural Questions · hits@5 | 68.8 -> 65.8 | No clear effect | Learning to Rank in Generative Retrieval (AAAI 2024 · 21 citations) |
| add the pre-trained font codebook | UFUC · SSIM | 0.793 -> 0.810 | Helped | VQ-FONT: Few-Shot Font Generation with Structure-Aware Enhancement and Quantization (AAAI 2024 · 21 citations) |
| Add input perturbation | CIFAR-100 · top-1 accuracy | 63.39 -> 64.38 | Helped | Robustness-Guided Image Synthesis for Data-Free Quantization (AAAI 2024 · 9 citations) |
| add diffusion-style noise during training | Cue-conflict images · shape bias | 21% -> 57% | Helped | Intriguing Properties of Generative Classifiers (ICLR 2024 · 9 citations) |
| Set the noise-input l2 distance to 0.5 rather than 0.3 | Average across CASIA, PolyU, HFUT, and TongJi · TAR@1e-4 | 0.9057 -> 0.9239 | Helped | PVTree: Realistic and Controllable Palm Vein Generation for Recognition Tasks (AAAI 2025 · 8 citations) |
| Use AutoPresent with SlidesLib instead of AutoPresent without SlidesLib | Detailed instructions with images · overall score | 45.2 -> 55.0 | Helped | AutoPresent: Designing Structured Visuals from Scratch (CVPR 2025 · 8 citations) |
| add the metrical controller | Sonnet test · semantic average | 61.70 -> 63.60 | Helped | PoetryDiffusion: Towards Joint Semantic and Metrical Manipulation in Poetry Generation (AAAI 2024 · 8 citations) |
| use 3D C2A2 instead of 2D AV representation | AffectNet · FED | 20.347 -> 16.060 | Helped | A Unified and Interpretable Emotion Representation and Expression Generation (CVPR 2024 · 7 citations) |
| Use QDGS pretraining with ArcFace rather than ArcFace baseline alone | RFW dark-skinned · accuracy | 79.7±0.3 -> 82.8±0.2 | Helped | Quality-Diversity Generative Sampling for Learning with Synthetic Data (AAAI 2024 · 6 citations) |
| Add Stable Diffusion generator artifacts to retrieved real images using SDEdit | FGVC-Aircraft · downstream accuracy | Hurt | The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · 6 citations) | |
| Add Self-Debug Learning | ACPR test · Acc | 31.89 -> 31.95 | Helped | Less Is More: Adaptive Program Repair with Bug Localization and Preference Learning (AAAI 2025 · 5 citations) |
What reviewers will expect
Datasets and baselines named in the 24 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 · 5 of 24 papers
- ImageNet · 4 of 24 papers
- FFHQ · 3 of 24 papers
- CelebA · 3 of 24 papers
- CIFAR-100 · 2 of 24 papers
- ImageNet-1K · 2 of 24 papers
- Tiny-ImageNet · 2 of 24 papers
- ImageNet-100 · 2 of 24 papers
Baselines
- FID · 3 of 24 papers
- GIF · 2 of 24 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 79 papers; from OpenAlex and the papers' affiliations).
- Google · 6 papers
- Shanghai Jiao Tong University · 5 papers
- University of California, Berkeley · 5 papers
- National University of Singapore · 4 papers
- Massachusetts Institute of Technology · 4 papers
- Tencent · 4 papers
- Zhejiang University · 4 papers
- Harbin Institute of Technology · 4 papers
- Carnegie Mellon University · 4 papers
- Tsinghua 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
- Classical generative image synthesis (147 papers)
- Retrieval and matching with generative features (140 papers)
- Learning-based image quality assessment (62 papers)
Elsewhere in the topic
- High-resolution diffusion synthesis (140 papers)
- Generative latent space tooling (129 papers)
- Generalizable fake image detection (119 papers)
- Latent diffusion with learned tokenizers (114 papers)
- Compositional scene generation (105 papers)
- Next-scale and token autoregressive images (90 papers)
- Robust diffusion watermarking (79 papers)
- Neural text-to-speech synthesis (72 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).