Living research atlas · data through 2026-08
How do you make a small dataset train like a large one?
Dataset condensation (also called dataset distillation) synthesizes or selects a tiny training set, sometimes one image per class, on which a model learns nearly what it would learn from the full data. Since 2018 the field has gone through five different answers to the question above, and it has spread from images to image–text pairs, video, graphs, text and more.
This atlas organizes that literature by what is condensed and how, rather than as a list. Every paper is checked against its primary source before anything is written about it.
Start with the synthesis
The field in 2026 →
Where unimodal and multimodal condensation stand today: the competing directions, what each bet on and gave up, where they converged, and what the evidence does and does not support.
The map
Rows are the five paradigms, ordered by how tightly the learner is coupled to synthesis. Columns are settings, meaning what is being condensed. Each cell counts catalogued papers, and empty cells are open territory. Click a cell to open those papers.
| Paradigm ↓ · Setting → | Images | Vision–language | Audio–visual & omni | Video | Graphs | Text | Time series | Pre-training & transfer | Dense prediction | Other data |
|---|---|---|---|---|---|---|---|---|---|---|
| P0 Selection Coreset selection · Dataset quantization | 13 | · | · | 3 | 1 | · | · | 1 | 2 | 2 |
| P1 Bi-level performance matching Meta-learning through unrolled training · Kernel and closed-form inner solvers | 27 | 2 | · | · | 2 | 1 | · | 2 | · | 8 |
| P2 Surrogate matching Gradient matching · Trajectory matching · Distribution and feature matching | 62 | 8 | 4 | 3 | 11 | 2 | 5 | 2 | 1 | 7 |
| P3 Decoupled synthesis Decoupled teacher-driven synthesis | 24 | · | · | 3 | · | · | · | 1 | 2 | · |
| P4 Generative priors Optimization in a generative latent space · The generator as the distilled artifact · Diffusion-based synthesis · Autoregressive and flow-based synthesis | 49 | 2 | · | 2 | · | 1 | · | 1 | 3 | 3 |
A paper is counted in every paradigm its families belong to. Design choices that attach to any paradigm (parameterization, soft labels,optimization recipes) are not rows.
Five answers, in the order they appeared
P0
Selection
Which real samples should be kept?
22 papers
P1
Bi-level performance matching
Does a model trained on the synthetic set perform like one trained on the real set?
41 papers
P2
Surrogate matching
Does the synthetic set produce the same gradients, trajectories, or features?
109 papers
P3
Decoupled synthesis
Can a frozen teacher alone tell us what the synthetic set should contain?
30 papers
P4
Generative priors
Can a pretrained generator supply the space, the artifact, or the sampler?
59 papers
Start here
New to the field →
The four questions every paper answers, and why the atlas is organized around them.
Multimodal condensation →
Image–text pairs: what changes when there are no classes and the learner is contrastive.
Generative condensation →
Why pretrained diffusion models became the fastest-growing family after 2024.
Reading the numbers →
Soft labels, evaluation protocols and budgets: why most leaderboards compare different things.
Newest methods
All papers →SRG — Self-Supervised Representation-Guided Generative Dataset Distillation
Mingzhuo Li, Guang Li, Linfeng Ye et al. · arXiv 2026Diffusion-based synthesisPre-training & transferpaper ↗
ProtoBlend — Efficient Video Dataset Distillation via Cluster-Guided Prototype Blending
Chongle Ren, Guang Li, Wenbo Huang et al. · arXiv 2026Dataset quantizationDecoupled teacher-driven synthesisVideopaper ↗
InfMatch — Dataset Distillation by Influence Matching
Haoru Tan, Wang Wang, Sitong Wu et al. · CVPR 2026Meta-learning through unrolled trainingOptimization and training recipesVision–languagepaper ↗code ↗
Adaptive Latent Trajectory Anchoring for Action Segmentation Dataset Condensation
Artheme Gauthier-Villar, Guodong Ding, Angela Yao · ECCV 2026Diffusion-based synthesisVideopaper ↗
CIM — Condensing Large-Scale Datasets Directly with Minimal Information Loss
Xinyi Shang, Peng Sun, Bei Shi et al. · ECCV 2026notableDecoupled teacher-driven synthesispaper ↗code ↗
RAHA — Rank-Aware Hyperbolic Alignment for Vision-Language Dataset Distillation
Jongoh Jeong, Sun-Kyung Lee, Kuk-Jin Yoon · ECCV 2026notableDistribution and feature matchingVision–languagepaper ↗code ↗
GADC — Geometry-Aware Dataset Condensation for Diffusion Model Training
Xiao Cui, Yulei Qin, Mo Zhu et al. · ICML 2026Coreset selectionpaper ↗code ↗
FOSTER — FOSTER: First-order Dataset Distillation for Text-based Sequential Recommendation
Hung Vinh Tran, Tong Chen, Xinyi Gao et al. · arXiv 2026Other datapaper ↗