Dataset Condensation Atlas

Reading the numbers

Evaluation

The atlas does not publish a leaderboard. The reason is not squeamishness: the variables that change between papers (labels, evaluation network, epochs, augmentation, budget accounting, backbone protocol) move accuracy by more than most method contributions do.

The standard protocol, and what it was meant to measure

In the class-conditional image setting, a distilled set S\mathcal{S} is evaluated by training a freshly initialized network on S\mathcal{S} alone and reporting top-1 accuracy on the real test set, averaged over several runs. The budget is counted in images per class (IPC). Small benchmarks (MNIST, CIFAR-10/100, Tiny-ImageNet) conventionally use a 3–4 layer ConvNet with the differentiable augmentation of DSA. ImageNet-1K uses ResNet-18/50.

The protocol is meant to measure the information in the synthetic data. Since 2023 its evaluation phase has absorbed more and more machinery that also carries information: a pretrained teacher that relabels every augmented crop in every epoch, strong augmentation such as CutMix, long schedules, and tuned batch sizes and learning-rate curves. Everything on this page comes from papers read in full.

Three label regimes, one set of images

Separate three regimes when reading any number:

Hard truths about soft labels evaluates the same subsets under all three. On ImageNet-1K at 50 IPC with ResNet-18 (Table 1):

ImagesHard labelsFixed soft labelsRelabel with KD
SRe2L synthetic images9.8%27.6%52.6%
Random real images28.5%48.5%58.1%
Full training set (reference)62.5%

Under hard labels the synthetic images trail random real images by about 19 points. Under relabel-with-KD the gap shrinks to about 5.5 points, and every method and coreset tested falls inside a 52–59% band. Once the teacher supervises every crop, the quality of the images barely matters.

The same effect appears inside individual papers. NRR-DD re-evaluates RDED’s own images: 42.0% with soft labels, 16.3% with one-hot labels (ImageNet-1K, 10 IPC, ResNet-18). The images are identical; the 25.7-point difference is the labels. CIM removes relabeling for several methods at the same setting: SRe2L falls to 1.1% and G-VBSM to 0.8%, while RDED keeps 19.7% and CIM 22.0%. Methods that stay close to real images survive the removal; methods that invert a teacher’s batch-norm statistics do not. A label is worth a thousand images had already shown that random real images given the same soft labels match or beat the synthesized images of state-of-the-art methods at higher IPC.

The recipe moves the same method by 20 points

The table below shows one nominal cell (SRe2L, ImageNet-1K, 10 IPC, ResNet-18, relabel with KD), as reported by different papers that re-ran it:

SourceSRe2L accuracy
SRe2L, original paper21.3%
CIM, re-run31.1%
DELT, re-run with CDA’s configuration41.9%
CV-DD, tuned “SRe2L++” baseline43.1%

That spread is larger than most improvements claimed in this family. RD³ makes it general: across eight decoupled and generative methods the published gap at 10 IPC is 27.3 points, and under one unified protocol (batch size and learning-rate schedule) it shrinks to 6.7 points. Every method’s own accuracy rises under the unified recipe, which means the original papers were under-tuned, not only incomparable. GIFT shows that the optimizer and loss used with soft labels can swing the same images and labels from 1.9% to 53.4% (RDED, 10 IPC, SGD without and with its loss). DC-BENCH found the same pattern in 2022, before soft labels: augmentation alone moved CIFAR-10 accuracy at 10 IPC by 6.6–10.2 points depending on the method.

Against random images and strong coresets

DD-Ranking scores each method against random real images trained with that method’s own recipe. On CIFAR-10, SRe2L’s improvement over random is negative at 1, 10 and 50 IPC, and RDED’s is slightly negative too. Hard-label trajectory matching scores best (+16.7 at 10 IPC). On ImageNet-1K with relabeling, current methods consistently underperform random selection given the same labels.

Do distilled sets outperform coresets? compares seven recent diffusion-based and decoupled methods against two diversity-aware coresets under three label protocols. On ImageNet-1K with hard labels and ResNet-18, the AUM coreset beats the best distilled set at 10 IPC (18.9% vs 15.6%) and at 50 IPC (43.5% vs 39.3%). With soft labels the gap narrows, but the coresets stay slightly ahead. The reversal comes on ImageNette, a small and easy dataset, where diffusion-based synthesis wins even under hard labels (60.0% vs 57.0% at 10 IPC). Coresets also cost a fixed 3–6 GPU-hours to build, while synthesis cost grows with IPC (12.8 to 51.7 hours for ManifoldGD from 10 to 50 IPC).

Where synthesis still clearly helps

The evidence is not that condensation fails. It is that its advantage is regime-specific:

The label bill

IPC counts images, not what else is stored. LPLD reports soft labels exceeding the images by more than 30× in ImageNet condensation, and compresses them from 113 GB to 2.8 GB at 200 IPC. HeLlO lists RDED’s soft-label storage at 572 MB, 5.7 GB and 28.6 GB for 1, 10 and 50 IPC, and replaces it with a projector of about 0.8M parameters. That wins at 1 and 10 IPC but loses at 50 (52.2% vs 56.5%). NRR-DD stores two scalars per crop instead of a probability vector, recovering about 70% of the soft-vs-hard gap at roughly 500× less label storage. HALD inserts a storage-free hard-label phase between two soft-label phases, and at 285 MB it beats LPLD by 9 points.

Multimodal evaluation is split too

Image–text distillation reports image↔text Recall@K on Flickr30K and MS-COCO at a fixed number of pairs. Papers differ in which encoders are pretrained, which are trained during evaluation, and which text encoder is used (for example, frozen versus trainable BERT). Numbers from different recipes belong in separate columns; the vision–language setting page records the protocol of each method.

What the evidence recommends

  1. Report hard-label accuracy first, and fixed-soft and relabel-with-KD accuracy on the same images, so the labels’ contribution is visible.
  2. Include random real images and a diversity-aware coreset, both trained with exactly the same labels, augmentation, optimizer, batch size, schedule and epochs as the method. A hardest-examples-only coreset is not enough.
  3. Count everything stored (soft labels, decoders, generators, projectors) and report construction time, stating whether teacher or generator pre-training is excluded.
  4. Report accuracy on an architecture not used during synthesis, ideally a transformer as well as a CNN. Batch-norm inversion methods degrade sharply on ViTs (RD³).
  5. Treat cross-paper comparisons as provisional unless the baselines were re-run under one recipe.

The atlas records these facts on each paper where its source states them, and never ranks numbers produced under different answers to questions 1–3.

Benchmarks and evaluation studies 13

2026-06

Do distilled sets beat coresets? — Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets?

Trisha Mittal, Akshay Mehra, Joshua Kimball · arXiv 2026notablepaper ↗

Benchmarks seven state-of-the-art dataset-distillation methods against three coreset-selection strategies on ImageNet-1K, ImageNet-100 and ImageNette under three standardized evaluation protocols, finding that some distillation methods fail to beat random subsets and that the strongest ones are comparable to or worse than coresets on these large-scale datasets while costing substantially more to construct, with coresets giving better distributional coverage.

2026-01

Towards Data Quality-Aware Dataset Distillation in Bioimaging

Bárbara Capelo, Maria Russo, André Carreiro et al. · BIOIMAGING 2026paper ↗

A preliminary study of quality-aware dataset distillation for blood cell microscopy images, arguing that fidelity and diversity metrics should be assessed alongside standard downstream performance measures, and examining how different initialization strategies affect the fidelity and diversity of the resulting distilled data.

2025-09

RD3 — Rectified Decoupled Dataset Distillation: A Closer Look for Fair and Comprehensive Evaluation

Xinhao Zhong, Shuoyang Sun, Xulin Gu et al. · ICLR 2026notablepaper ↗code ↗

RD3 systematically studies how post-evaluation choices (augmentation, epoch-wise soft labels, and other protocol details introduced after decoupled distillation methods such as SRe2L) affect reported test accuracy, and finds that much of the performance variation between existing decoupled-distillation methods reflects inconsistent evaluation protocols rather than differences in the intrinsic quality of the synthetic data. The paper proposes a standardized benchmark and evaluation protocol, and identifies general strategies that improve distilled-dataset effectiveness across settings.

2025-05

DD-Ranking — DD-Ranking: Rethinking the Evaluation of Dataset Distillation

Zekai Li, Xinhao Zhong, Samir Khaki et al. · arXiv 2025corepaper ↗code ↗

Shows that reported accuracy gains from decoupled dataset-distillation pipelines largely come from soft labels and stronger post-hoc augmentation rather than the intrinsic quality of the distilled images themselves, since even randomly sampled images can match state-of-the-art distilled sets under the same labeling and augmentation recipe, and proposes DD-Ranking, a unified evaluation framework and set of metrics that isolate the images' own contribution.

2024-12

Video DC study — A Large-Scale Study on Video Action Dataset Condensation

Yang Chen, Sheng Guo, Bo Zheng et al. · arXiv 2024Videopaper ↗code ↗

This large-scale study systematically examines video dataset condensation along three axes: temporal processing of video data, the evaluation protocol, and how condensation algorithms adapt to the space-time domain. It finds that labeling method matters a great deal for condensation performance, that simple sliding-window sampling suffices for temporal processing, and that dataset-distillation methods win in harder scenarios while sample-selection methods win in easier ones; it also proposes a unified evaluation protocol and reports state-of-the-art results on HMDB51, UCF101, SSv2 and K400 under it.

2024-11

BEARD — BEARD: Benchmarking the Adversarial Robustness for Dataset Distillation

Zheng Zhou, Wenquan Feng, Shuchang Lyu et al. · arXiv 2024paper ↗code ↗

Introduces BEARD, an adversarial-robustness benchmark for dataset distillation covering DM, IDM and BACON distilled sets on CIFAR-10/100 and Tiny-ImageNet under FGSM/PGD/C&W attacks; defines a Robustness Ratio, an Attack Efficiency Ratio and a combined Comprehensive Robustness-Efficiency Index, and publishes results on a public leaderboard across several images-per-class settings and adversarial-training conditions.

2024-07

Dataset Distillation in Medical Imaging: A Feasibility Study

Muyang Li, Can Cui, Quan Liu et al. · arXiv 2024paper ↗

Empirically tests whether existing dataset distillation methods, developed for natural images, transfer to medical imaging, across datasets with minor and major domain variation, and asks whether a small representative sample can predict distillation success. Reports that distillation can substantially reduce medical dataset size while keeping performance comparable to full-data training, and that a small representative image sample is a reliable indicator of how well distillation will work on a given dataset.

2024-06

GC-Bench — GC-Bench: An Open and Unified Benchmark for Graph Condensation

Qingyun Sun, Ziying Chen, Beining Yang et al. · NeurIPS 2024notableGraphspaper ↗code ↗

Introduces GC-Bench, a benchmark that systematically evaluates 12 graph-condensation methods on node- and graph-level tasks across 12 datasets along effectiveness, transferability and complexity, and releases a reusable library for training and evaluating graph condensation methods.

2024-06

GC4NC — GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights

Shengbo Gong, Juntong Ni, Noveen Sachdeva et al. · NeurIPS 2025 Datasets and Benchmarks TrackGraphspaper ↗code ↗

Introduces GC4NC, a benchmark that evaluates graph-condensation methods for node classification across performance, efficiency, privacy preservation, denoising ability, NAS effectiveness and transferability, reporting new insights into which design choices drive condensed-graph quality along these dimensions.

2024-05

GCondenser — GCondenser: Benchmarking Graph Condensation

Yilun Liu, Ruihong Qiu, Zi Huang · arXiv 2024Graphspaper ↗code ↗

Introduces GCondenser, the first large-scale graph-condensation benchmark with a standardized condensation-validation-evaluation pipeline that extends to new methods and datasets, and uses it to comprehensively compare mainstream graph-condensation methods.

2024-03

DD-RobustBench — DD-RobustBench: An Adversarial Robustness Benchmark for Dataset Distillation

Yifan Wu, Jiawei Du, Ping Liu et al. · TIP 2025paper ↗code ↗

Extends adversarial-robustness benchmarking of distilled datasets to a wider range of methods (including TESLA and SRe2L) and to ImageNet-1K, evaluating against PGD and AutoAttack and analyzing robustness from a frequency perspective; finds that mixing distilled data into standard training batches on the original dataset can improve robustness.

2023-05

A Comprehensive Study on Dataset Distillation: Performance, Privacy, Robustness and Fairness

Zongxiong Chen, Jiahui Geng, Derui Zhu et al. · arXiv 2023paper ↗

Runs membership-inference, robustness, and per-class fairness evaluations across several dataset-distillation methods and reports that distillation reduces but does not eliminate privacy leakage, can degrade adversarial robustness to varying degrees, and can amplify unfairness across classes; proposes a large-scale benchmarking framework covering these axes.

2022-07

DC-BENCH — DC-BENCH: Dataset Condensation Benchmark

Justin Cui, Ruochen Wang, Si Si et al. · NeurIPS 2022corepaper ↗code ↗

Introduces DC-BENCH, the first large-scale standardized benchmark for dataset condensation, evaluating methods under controlled data-augmentation and architecture choices so that gains attributable to the condensation method itself can be separated from evaluation-pipeline confounds, and releases evaluators, baselines and generated datasets for reproducible comparison.

Analysis and theory 20

2026-06

Structural Assessment for Understanding and Guiding Dataset Distillation in Discrete Token Space

Yue Cao, Jianyang Gu, Vyacheslav Kungurtsev et al. · ECCV 2026paper ↗

Analyzes distilled datasets through discrete visual tokenizers, introducing a structural score that measures how balanced a distilled dataset's token-level composition is, and finds that balanced token composition — not divergence from the original data distribution — correlates with higher validation performance; shows that samples with high structural scores can guide diffusion-based dataset distillation toward more effective synthetic sets.

2026-04

Hard truths about soft labels — Rethinking Dataset Distillation: Hard Truths about Soft Labels

Priyam Dey, Aditya Sahdev, Sunny Bhati et al. · CVPR 2026notablepaper ↗

Shows that with abundant or fixed soft labels from a pretrained teacher, coresets fail to convincingly beat random subsets and performance saturates near the full-dataset level regardless of subset quality — unlike the hard-label coreset literature — and that in the hard-label setting only RDED reliably beats random baselines on ImageNet-1K among nine distillation methods tested; introduces a compute-aware pruning metric (CAD-Prune) and a resulting method (CA2D) that outperforms existing distillation methods on ImageNet-1K under a fixed compute budget.

2026-03

Dataset Distillation Efficiently Encodes Low-Dimensional Representations from Gradient-Based Learning of Non-Linear Tasks

Yuri Kinoshita, Naoki Nishikawa, Taro Toyoizumi · ICML 2026paper ↗

Proves for gradient-based training of two-layer networks on a multi-index-model task that dataset distillation encodes the task's low-dimensional structure into the synthetic data, achieving a required memory complexity of order r^2 d + L (r = intrinsic dimension, d = input dimension, L = network width) — one of the first theoretical analyses tying the compression rate to a specific task structure for a gradient-based distillation algorithm.

2025-12

Utility boundary laws — Utility Boundary of Dataset Distillation: Scaling and Configuration-Coverage Laws

Zhengquan Luo, Zhiqiang Xu · ICML 2026paper ↗

Proposes a unified 'configuration-dynamics-error' framework that reformulates gradient-, distribution- and trajectory-matching dataset distillation as interchangeable surrogates reducing the same generalization error, deriving a scaling law bounding error against distilled sample size (explaining observed performance saturation) and a coverage law showing the required sample size grows linearly with training-configuration diversity, both with matching upper and lower bounds confirmed experimentally.

2025-09

A Discrepancy-Based Perspective on Dataset Condensation

Tong Chen, Raghavendra Selvan · arXiv 2025paper ↗

Recasts dataset condensation as minimizing a discrepancy between the real and synthetic data distributions, showing that gradient-, distribution- and trajectory-matching objectives are instances of one framework and extending the objective beyond test accuracy to robustness and privacy criteria.

2025-06

Flowing Datasets with Wasserstein over Wasserstein Gradient Flows

Clément Bonet, Christophe Vauthier, Anna Korba · ICML 2025paper ↗code ↗

Represents a labeled dataset as a mixture distribution over per-class feature distributions and defines gradient flows on this space of distributions-over-distributions using a Wasserstein-over-Wasserstein metric, applying the resulting dynamics with a sliced-Wasserstein-kernel maximum-mean-discrepancy objective to transfer-learning and dataset-distillation tasks.

2025-03

Spectral filtering view — Understanding Dataset Distillation via Spectral Filtering

Deyu Bo, Songhua Liu, Xinchao Wang · ICLR 2026paper ↗

Interprets every dataset-distillation objective as a filter function acting on the eigenvalues of the feature-feature correlation matrix and the frequencies of the feature-label correlation matrix, showing existing methods split into low-frequency (global texture) and high-frequency (local detail) matchers under one spectral-filtering framework, and proposes Curriculum Frequency Matching, which adjusts the filter during distillation to cover both regimes, outperforming baselines on CIFAR-10/100 and ImageNet-1K.

2024-09

Dataset Distillation from First Principles: Integrating Core Information Extraction and Purposeful Learning

Vyacheslav Kungurtsev, Yuanfang Peng, Jianyang Gu et al. · arXiv 2024paper ↗

Argues dataset distillation is under-specified without naming the downstream inference task, formalizes DD relative to a task-specific objective, and uses the formalization to analyze existing methods' faithfulness, with case studies on merging medical datasets that share only some features and on generating more physically faithful boundary data for physics-informed neural networks.

2024-06

A label is worth a thousand images — A Label is Worth a Thousand Images in Dataset Distillation

Tian Qin, Zhiwei Deng, David Alvarez-Melis · NeurIPS 2024landmarkpaper ↗code ↗

Ablates the many disparate dataset-distillation methods and finds that the single largest common factor in their reported performance is the use of soft (probabilistic) labels rather than their specific synthesis techniques; shows the soft labels must carry structured information to help, provides empirical scaling laws relating soft-label effectiveness to images-per-class, and establishes an empirical Pareto frontier for data-efficient learning, challenging the attribution of gains to image-synthesis method alone.

2024-06

What is DD learning? — What is Dataset Distillation Learning?

William Yang, Ye Zhu, Zhiwei Deng et al. · ICML 2024notablepaper ↗code ↗

Finds that distilled data cannot substitute for real data outside the standard dataset-distillation evaluation setting, that the distillation process retains high task performance mainly by compressing information about real models' early training dynamics, and provides an interpretation framework showing individual distilled data points carry meaningful semantic information.

2023-12

ELF — Boosting the Cross-Architecture Generalization of Dataset Distillation through an Empirical Study

Lirui Zhao, Yuxin Zhang, Fei Chao et al. · arXiv 2023paper ↗code ↗

Attributes the poor cross-architecture transfer of distilled datasets to an inductive bias the synthetic images inherit from the network used to distill them, then proposes ELF, an evaluation procedure that trains the evaluation network against intermediate features of the distillation model instead of only labels. Reports that ELF improves cross-architecture accuracy for several existing distillation methods without changing the distilled images themselves.

2023-12

A Theoretical Study of Dataset Distillation

Zachary Izzo, James Zou · NeurIPS 2023 Workshoppaper ↗

Proves existence and impossibility results for exact dataset distillation of generalized linear models: a single synthetic point can exactly reproduce a model trained on the full data for GLMs, and a size-independent construction exists for linear regression with any data-independent regularizer, but no analogous construction exists for logistic regression, and kernel regression cannot in general be distilled to recover even one model exactly.

2023-10

Can pre-trained models assist in dataset distillation?

Yao Lu, Xuguang Chen, Yuchen Zhang et al. · arXiv 2023paper ↗code ↗

Empirically studies which properties of a pretrained model used to guide dataset distillation matter, finding that greater model diversity helps, that sub-optimal (not fully converged) models can outperform well-trained ones, and that only a reasonable domain match — not an exact one — is needed, then uses these findings to improve cross-architecture generalization over baseline distillation methods.

2023-05

Size & approximation error — On the Size and Approximation Error of Distilled Sets

Alaa Maalouf, Murad Tukan, Noel Loo et al. · NeurIPS 2023paper ↗

Gives the first proof, for kernel-ridge-regression-based distillation methods such as Kernel Inducing Points, that a small distilled set exists for shift-invariant kernels by transforming the ridge-regression problem into random Fourier feature space, showing the distilled-set size can be linear in the RFF-space dimension (or near-linear in the kernel's effective degrees of freedom) with an excess-risk bound that depends on the regularization parameter.

2022-10

On Divergence Measures for Bayesian Pseudocoresets

Balhae Kim, Jungwon Choi, Seanie Lee et al. · NeurIPS 2022paper ↗code ↗

Casts two representative dataset-distillation algorithms as approximations to Bayesian-pseudocoreset construction that minimize reverse KL divergence and Wasserstein distance respectively, giving a unifying view of divergence measures for pseudocoreset construction, and proposes a new pseudocoreset algorithm that instead minimizes forward KL divergence; shows empirically that pseudocoresets built this way better reflect the true posterior even in high-dimensional Bayesian inference problems.

2022-09

No Free Lunch — No Free Lunch in "Privacy for Free: How does Dataset Condensation Help Privacy"

Nicholas Carlini, Vitaly Feldman, Milad Nasr · arXiv 2022notablepaper ↗

A rebuttal of Dong et al. (2022, "Privacy for Free"), arguing its empirical evaluation and theoretical analysis of dataset condensation's privacy benefit both have major flaws, so the original paper does not provide statistically significant evidence that DC improves privacy over a naive baseline. Points out that DP-SGD, the standard privacy-preserving training approach, simultaneously gives better accuracy and a provably lower membership-attack success rate than the condensation-based claims it disputes.

2022-07

On Implicit Bias in Overparameterized Bilevel Optimization

Paul Vicol, Jonathan P. Lorraine, Fabian Pedregosa et al. · ICML 2022paper ↗

Studies how the algorithmic choices of bi-level optimization — cold-start versus warm-start inner-loop initialization, and the hypergradient approximation used — implicitly bias which solution gradient-based bi-level methods converge to, in settings including hyperparameter optimization, meta-learning and dataset distillation, and shows warm-start solutions can retain substantial information about the outer objective even in low-dimensional problems.

2020-10

New Properties of the Data Distillation Method When Working With Tabular Data

Dmitry Medvedev, Alexander D'yakonov · AIST 2020Other datapaper ↗code ↗

This paper applies the original backpropagation-through-time dataset distillation method to tabular data and finds that a model trained on the distilled samples can outperform one trained on the original dataset, but that the distilled data generalizes poorly to models with different hyperparameters than the one used during distillation. It shows that distilling with multiple architectures simultaneously mitigates this poor cross-hyperparameter generalization.

2019-11

Optimizing Millions of Hyperparameters by Implicit Differentiation

Jonathan Lorraine, Paul Vicol, David Duvenaud · AISTATS 2020paper ↗code ↗

Proposes an implicit-function-theorem approach to hyperparameter gradients using efficient inverse-Hessian approximations, enabling joint optimization of millions of weights and hyperparameters (including a learned data-augmentation network) at only a few times the cost of standard training — an implicit-differentiation technique later adopted by kernel-based dataset distillation methods to avoid unrolling the inner loop.

2015-02

Hypergradient — Gradient-based Hyperparameter Optimization through Reversible Learning

Dougal Maclaurin, David Duvenaud, Ryan P. Adams · ICML 2015paper ↗code ↗

Computes exact gradients of validation performance with respect to thousands of hyperparameters — including a data-augmentation network whose weights are treated as hyperparameters — by exactly reversing SGD-with-momentum dynamics, a hypergradient technique later reused to differentiate through the training procedure in bi-level dataset distillation.