Dataset Condensation Atlas

How this atlas is organized

Four questions, not one tree

A paper on trajectory matching for image–text pairs with a factorized representation belongs under three headings in a conventional list, and loses two of them. The atlas instead places every paper by answering four questions separately. Each answer is a facet you can filter on.

Q1

Setting

What is being condensed, for which learner?

10 settings

Q2

Paradigm & family

How is the synthetic set obtained?

5 paradigms

Q3

Design choices

What is stored, what is the label, how is it optimized?

3 orthogonal groups

Q4

Lens

What kind of contribution is it?

7 kinds

Q2 in one picture: paradigms ordered by coupling

Moving down the ladder, less of the learner sits inside the synthesis loop. Cost fell and scale rose in the same order, so the ladder doubles as a history of the field.

P0

Selection

Coupling: none — no synthesis

Which real samples should be kept? The distilled set is a subset of the real data, chosen by a scoring or covering rule.

Coreset selection 14 · Dataset quantization 8

P1

Bi-level performance matching

Coupling: the full inner training loop is differentiated (or solved analytically)

Does a model trained on the synthetic set perform like one trained on the real set? Optimize the synthetic set through the training of a model on it, by unrolling or in closed form.

Meta-learning through unrolled training 21 · Kernel and closed-form inner solvers 20

P2

Surrogate matching

Coupling: one network (or a short segment of its training) per update

Does the synthetic set produce the same gradients, trajectories, or features? Replace the bi-level objective with a single-level proxy computed through networks in the loop.

Gradient matching 26 · Trajectory matching 33 · Distribution and feature matching 50

P3

Decoupled synthesis

Coupling: a frozen pretrained teacher; no student in the synthesis loop

Can a frozen teacher alone tell us what the synthetic set should contain? Train a teacher once, synthesize or select from it without any bi-level coupling, then train students on its soft labels.

Decoupled teacher-driven synthesis 30

P4

Generative priors

Coupling: a frozen (or lightly fine-tuned) generator; often no learner in the loop at all

Can a pretrained generator supply the space, the artifact, or the sampler? Use a pretrained generative model to parameterize, store, or directly sample the synthetic set.

Optimization in a generative latent space 9 · The generator as the distilled artifact 11 · Diffusion-based synthesis 36 · Autoregressive and flow-based synthesis 3

+

Orthogonal design choices

Coupling: independent of paradigm

What is stored, what is the label, and how is it optimized? Choices that attach to a method in any paradigm — parameterization, label handling, optimization recipes.

Synthetic-set parameterization 28 · Label distillation and soft labels 20 · Optimization and training recipes 46

Why not an existing organization?

Seven surveys and the community index were read in full before this taxonomy was fixed (notes in the repository under docs/recon/surveys.md). They organize the field along different axes, and they disagree on basic placements.

SurveyOrganizing axisCovers decoupled, soft-label, generative, multimodal work?
Yu, Liu & Wang, TPAMI 2023Optimization objective: performance, parameter, distribution matchingNo
Lei & Tao, TPAMIMeta-learning versus data-matching frameworks; factorized variants marked across the treeNo
Sachdeva & McAuley, TMLR 2023Approach family, then data modalityNo
Geng et al., IJCAI 2023Learning frameworks plus common enhancements; rich applicationsNo
Liu & Du, 2025Matching, scalable/generative, decoupledYes: the only one with first-class sections for all four
Gao et al., TKDE 2025Graphs only; by evaluation criterionGraphs only
Hashemi et al., IJCAI 2024Graphs only; condensation as a sibling of sparsification and coarseningGraphs only

Where they disagree.

The Awesome list’s headings, mapped

The Awesome Dataset Distillation list is the discovery source for this atlas, and its headings are a good index. They are not a classification, because they mix kinds of thing:

Awesome headingKind of thingWhere it goes in the atlas
Gradient/Trajectory Matching, Distribution/Feature Matching, Kernel-BasedMechanismQ2 families in P1 and P2
Decoupled DistillationMechanismQ2: P3 decoupled synthesis
Generative Distillation (GAN / Diffusion / VAR / Flow)MechanismQ2: P4 families
Distilled Dataset ParametrizationWhat is storedQ3 parameterization
Label DistillationWhat the label isQ3 label distillation
Better OptimizationHow it is optimizedQ3 optimization recipes, or the family it improves
Dataset QuantizationMechanism (selection)Q2: P0
Multimodal, Video, Graph, Text, Time Series, …What dataQ1 settings
Better Understanding, Benchmark, SurveyContribution typeQ4 lens
Privacy, Robustness, Fairness, Continual, Federated, Medical, …Contribution type / useQ4 lens, plus the mechanism family used

Two meanings of “multimodal”

The word is used for two different things, and related-work sections often mix them.

  1. The distilled dataset is multimodal. Image–caption pairs are condensed so that a vision–language model can be trained on them. This is the vision–language setting.
  2. A multimodal model is a tool. A text-to-image diffusion model, CLIP text prototypes or a vision–language model is used to condense an ordinary image classification dataset. These papers are in the image-classification setting, with a vision–language prior recorded as a design facet.

Why the paradigms are ordered by coupling

The ladder P1 → P4 removes the learner from the synthesis loop one piece at a time. P1 differentiates through training itself. P2 keeps one network in the loop but replaces training with a proxy (gradients, trajectory segments, feature distributions). P3 keeps only a frozen teacher. P4 keeps only a frozen generator and often no classifier at all. Cost fell and achievable scale rose in the same order, which is why the order also reads as a history of the field. It also points to where the field’s current weakness comes from: the less the learner is involved in synthesis, the more the evaluation-time teacher has to supply. See Evaluation.