Dataset Condensation Atlas

+ · Orthogonal design choices · since 2020 · contested

Label distillation and soft labels

Treat labels as part of what is distilled — learned labels, soft teacher labels, relabeling at every epoch — and account for what that costs.

The bet every member shares

A label vector can carry far more information about the dataset than a one-hot class index, and this information is cheap relative to images.

Defining objective

Lstudent=ExS,A  KL(σ(fteacher(A(x))/τ)σ(fstudent(A(x))/τ))\mathcal{L}_{\text{student}}=\mathbb{E}_{x\in\mathcal{S},\,A}\;\mathrm{KL}\Bigl(\sigma\bigl(f_{\text{teacher}}(A(x))/\tau\bigr)\,\Big\|\,\sigma\bigl(f_{\text{student}}(A(x))/\tau\bigr)\Bigr)

Relabeling with a pretrained teacher under random augmentation A; earlier work instead learns a fixed label vector jointly with the images.

Labels were always part of the dataset

A distilled image with a one-hot label carries only what its pixels carry. Give it a probability vector over classes and it also carries how similar the classes are, which is information the pixels would otherwise have to encode. Learned labels made this explicit in 2020 by distilling labels for real images instead of distilling images. Kernel methods learned labels alongside their support sets, and TESLA introduced soft labels to help trajectory matching converge on ImageNet-1K’s 1,000 classes. The decisive step was SRe2L’s relabel stage. A pretrained teacher labels every augmented crop of the synthetic images in every epoch, and students learn from those labels. Since 2023 almost every ImageNet-scale headline number has used this regime.

How much of the accuracy is the label?

Separate three regimes: hard labels, fixed soft labels (one stored vector per image), and relabel with knowledge distillation (a teacher labels every crop, every epoch). Papers read in full give consistent answers.

With a teacher supervising every crop, much of what reported accuracy measures is the teacher.

The label bill, and how to shrink it

Soft labels for every crop of every image take space that images-per-class does not count.

PaperApproachResult
LPLDBatch synthesis within classes so images are more diverse, then prune labels at randomImageNet-1K labels at 200 IPC from 113 GB to 2.8 GB; 40×-pruned labels still beat unpruned SRe2L (59.6% vs 57.0%)
HeLlOReplace stored labels with a CLIP-based image-to-label projector (~0.8M parameters)Beats RDED at 1 and 10 IPC but loses at 50 (52.2% vs 56.5%); RDED’s labels are 28.6 GB at 50 IPC
INFEROne static label per instance on real images plus a learned compensatorUp to 99% less label storage
NRR-DDStore two scalar distances per crop instead of a vectorAbout 70% of the soft-vs-hard gap recovered at roughly 500× less storage
HALDInsert a storage-free, label-smoothed hard-label phase between soft phases42.7% vs LPLD’s 33.7% at 285 MB (ImageNet-1K, 50 IPC)

The pattern is clear. Most of what the soft labels carry can be kept at a small fraction of their storage, but every method still depends on a teacher trained on the full dataset.

The training recipe around labels is itself a variable

GIFT shows that how soft labels are consumed matters as much as having them. With the same images and labels, changing only the optimizer and loss moves RDED at 10 IPC from 1.9% (SGD, standard loss) to 53.4% (SGD with GIFT’s cosine loss). A reported accuracy can therefore depend on an undisclosed optimizer–loss pairing more than on the data.

Risks

Soft labels can leak held-out teacher knowledge: in controlled settings, from toy models up to GPT-2, students trained on a memorizing teacher’s soft labels recover information about training examples they never saw. Higher softmax temperature increases the leakage. This has not been measured on a released condensation dataset. But teacher-relabeled “distilled datasets” should not be assumed to be private, or independent of the teacher.

Where the family stands

Labels are no longer a detail of evaluation; they are half of what is distilled. The evidence favors three practices: report hard-label accuracy alongside soft-label accuracy on the same images; count label storage in the budget; and compare against random real images given the same labels. See Evaluation and the open problems on uncounted label storage and teacher knowledge transfer.

Branches

Learned labels

Optimize label vectors jointly with (or instead of) the synthetic inputs.

Learned labels

Teacher relabeling and its cost

Soft labels from a pretrained teacher, and methods that prune, quantize or lighten the label storage they require.

LPLD, HeLlO, GIFT

Critiques

Analyses showing how much of the reported accuracy the labels, rather than the images, account for.

A label is worth a thousand images

How the family developed

The papers that moved the family, in the role each one played.

Origin

Learned labels · NeurIPS 2020 Workshop

Distills labels for a small set of real images instead of distilling images, and reports better compatibility with off-the-shelf optimizers and architectures than image-based distillation.

Improvement

LPLD · NeurIPS 2024

Batching synthesis within classes raises within-class diversity so that soft labels can be pruned at random: ImageNet-1K soft labels at 200 IPC shrink from 113 GB to 2.8 GB.

HeLlO · ICCV 2025

Replaces stored soft labels with a CLIP-based image-to-label projector adapted by low-rank fine-tuning, at about 0.003% of the soft-label storage.

GIFT · ICLR 2025

Finds students highly sensitive to the loss applied to soft labels and adds label refinement with a cosine-similarity loss, notably improving generalization across optimizers.

Challenge

A label is worth a thousand images · NeurIPS 2024

Ablations find that soft labels, not the image-synthesis technique, are the main factor behind state-of-the-art accuracy, and that only soft labels carrying structured information help.

What it gets right

  • Large accuracy gains in essentially every paradigm, particularly at ImageNet scale.

Where it is weak

  • Per-epoch soft labels for many augmented crops can require more storage than the images they label.
  • Makes comparisons with hard-label methods invalid unless reported separately; a teacher trained on the full data leaks its knowledge into the student.

Applied to

Papers per year

19
20
23
24
25
26

All papers in this family 20

Papers not already discussed above, ordered by tier, then newest first.Open in the explorer →

2024-06

LoRS — Low-Rank Similarity Mining for Multimodal Dataset Distillation

Yue Xu, Zhilin Lin, Yusong Qiu et al. · ICML 2024coreVision–languagepaper ↗code ↗

LoRS distills a ground-truth image-text similarity matrix alongside the synthetic pairs, instead of assuming one caption strictly matches one image, and stores that similarity structure efficiently via low-rank factorization. The paper reports significant improvements over prior vision-language distillation algorithms and proposes LoRS as a foundational synthetic-data setup for image-text dataset distillation.

2026-04

Soft Label Pruning and Quantization for Large-Scale Dataset Distillation

Xiao Lingao, Yang He · TPAMI 2026notablepaper ↗code ↗

Extends the within-class-batching idea for reducing soft-label storage with two further mechanisms: Label Pruning with Dynamic Knowledge Reuse improves label diversity per augmentation, and Label Quantization with Calibrated Student-Teacher Alignment improves diversity per image; reduces soft-label storage by 78x on ImageNet-1K and 500x on ImageNet-21K while improving accuracy by up to 7.2 and 2.8 points respectively.

2025-12

HALD — Hard Labels In! Rethinking the Role of Hard Labels in Mitigating Local Semantic Drift

Jiacheng Cui, Bingkui Tong, Xinyue Bi et al. · ICML 2026notablepaper ↗code ↗

Shows theoretically and empirically that limiting soft-label crops per image to save storage causes local semantic drift -- visually ambiguous crops pull soft supervision away from the image's true label -- and that hybridizing hard labels back in as a content-invariant semantic anchor corrects it; HALD uses hard labels as an intermediate corrective signal alongside sparse soft labels, reaching 42.7% top-1 on ImageNet-1K with only 285MB of soft-label storage (a 100x reduction) and a 9.0-point gain over LPLD.

2025-11

Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation

Xiao Cui, Yulei Qin, Wengang Zhou et al. · NeurIPS 2025notablepaper ↗

Reformulates generative dataset distillation as Optimal Transport distance minimization to capture instance-level and intra-class geometry that global mean/variance matching misses, with three components: OT-guided diffusion sampling that aligns latent distributions, label-image-aligned soft relabeling that adapts label distributions to the complexity of the distilled images, and OT-based logit matching between student outputs and soft-label distributions; reports at least a 4% accuracy gain over the state of the art at IPC=10 on ImageNet-1K across architectures.

2025-11

RLDD — Rethinking Long-tailed Dataset Distillation: A Uni-Level Framework with Unbiased Recovery and Relabeling

Xiao Cui, Yulei Qin, Xinyue Li et al. · AAAI 2026notablepaper ↗code ↗

Moves long-tailed dataset distillation away from trajectory matching to a statistical-alignment perspective, recalibrating Batch Normalization statistics via a full forward pass with dynamically adjusted momentum, using separate observer and teacher expert models for reliable statistics and soft-label generation, and initializing synthetic images by incrementally selecting high-confidence, diverse augmentations. Reports 15.6% top-1 accuracy improvement on CIFAR-100-LT and 11.8% on Tiny-ImageNet-LT at IPC=10, IF=10, over prior state-of-the-art long-tailed distillation methods.

2025-11

ADSA — Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation

Chenyang Jiang, Hang Zhao, Xinyu Zhang et al. · NeurIPS 2025notablepaper ↗code ↗

Derives an imbalance-aware generalization bound for models trained on distilled data and traces long-tailed dataset distillation's performance loss to soft-label bias entangled from two sources, the distillation model and the distilled images themselves, identified through systematic perturbation of imbalance levels; proposes ADSA, a lightweight Adaptive Soft-label Alignment module that calibrates this bias and plugs into existing distillation pipelines. Reports up to 11.8% tail-class accuracy improvement and 41.4% overall accuracy on ImageNet-1k-LT with EDC at IPC=50.

2024-08

INFER — Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensator

Xin Zhang, Jiawei Du, Ping Liu et al. · ICLR 2025notablepaper ↗code ↗

Breaks the one-hot, per-class synthesis barrier of prior distillation with a Universal Feature Compensator (UFC) that enhances feature integration across classes and can generate multiple additional synthetic instances from a single UFC input, plus label interpolation across classes that shrinks stored soft-label size toward zero; reports outperforming SRe2L by 34.5% at IPC=50 on ImageNet-1K with ResNet-18.

2024-05

ReLA — Efficiency for Free: Ideal Data Are Transportable Representations

Peng Sun, Yi Jiang, Tao Lin · NeurIPS 2024notablePre-training & transferpaper ↗code ↗

ReLA (Representation Learning Accelerator) argues that a publicly available, task- and architecture-agnostic "prior model" can be used to produce efficient training signal that is transportable across tasks and architectures, substituting learned representation targets for the labels/pipeline a representation-learning method would otherwise need. The paper reports that using a CIFAR-10-pretrained ResNet-18 as the prior model to inform ResNet-50 training on ImageNet-1K cuts computational cost by 50% while matching the accuracy of the original BYOL recipe trained at full cost.

2026-03

VQAE — Vector-Quantized Soft Label Compression for Dataset Distillation

Ali Abbasi, Ashkan Shahbazi, Hamed Pirsiavash et al. · arXiv 2026Textpaper ↗

Quantifies the bit cost of soft teacher labels across augmentations in dataset distillation, showing they dominate total storage in large-class settings such as ImageNet-1K, and introduces a vector-quantized autoencoder (VQAE) to compress them. Reports 30-40x additional compression over RDED, LPLD, SRe2L and CDA on ImageNet-1K while retaining over 90% of their original accuracy, and validates the approach on both vision and language distillation benchmarks.

2025-06

Soft labels leak — Dataset distillation for memorized data: Soft labels can leak held-out teacher knowledge

Freya Behrens, Lenka Zdeborová · ICLR 2026paper ↗

Shows that students trained on a teacher's soft labels can achieve non-trivial, sometimes near-perfect, accuracy on held-out data the teacher memorized but never generalized to, including on i.i.d. random datasets where generalization is a priori impossible, and that with enough soft labels the student can recover the teacher's predictions on all inputs, including the memorized held-out ones; the effect depends strongly on the softmax temperature used for the labels but persists across network capacities, architectures and dataset compositions.

2025-05

Data-Distill-Net — Data-Distill-Net: A Data Distillation Approach Tailored for Reply-based Continual Learning

Wenyang Liao, Quanziang Wang, Yichen Wu et al. · arXiv 2025paper ↗

For replay-based continual learning, keeps a learnable memory buffer of real replay images but, instead of optimizing the images themselves, trains a lightweight module that generates learned soft labels distilling both the current task's data and knowledge accumulated in the prior buffer. Reports competitive accuracy and reduced forgetting across continual-learning benchmarks relative to standard replay baselines, at lower overhead than distilling the buffer images directly.

2024-10

DRUPI — DRUPI: Dataset Reduction Using Privileged Information

Shaobo Wang, Youxin Jiang, Tianle Niu et al. · arXiv 2024paper ↗

Introduces Dataset Condensation using Privileged Information (DRUPI/DCPI): alongside condensed images and labels, synthesizes auxiliary feature or attention labels as an additional training target; finds that moderately (not maximally) discriminative and diverse feature labels work best, and shows the technique plugs into existing condensation methods for consistent gains on ImageNet-1K, CIFAR-10/100 and Tiny-ImageNet.

2024-09

Label-Augmented Dataset Distillation

Seoungyoon Kang, Youngsun Lim, Hyunjung Shim · WACV 2025paper ↗

Sub-samples each synthetic image to generate additional dense labels that capture richer semantics than one label per image, at only about 2.5% extra storage on ImageNet subsets; the label-generation strategy is complementary to existing distillation methods, improving three high-performance algorithms by an average of 14.9% accuracy and improving cross-architecture robustness.

2023-01

Dataset Distillation with Attention Labels for Fine-tuning BERT

Aru Maekawa, Naoki Kobayashi, Kotaro Funakoshi et al. · ACL 2023Textpaper ↗code ↗

Proposes an attention-label distillation method for fine-tuning BERT on condensed text data, transferring attention-probability supervision from a teacher alongside a small number of learned synthetic examples in a bi-level, few-gradient-step distillation setup; reports 98.5% of full-dataset performance from a single example per class and a single gradient step, including 93.2% accuracy on AG News.

2019-10

Soft-label DD — Soft-Label Dataset Distillation and Text Dataset Distillation

Ilia Sucholutsky, Matthias Schonlau · IJCNN 2021Textpaper ↗code ↗

Extends dataset distillation with soft (distributional) labels learned jointly with the synthetic images instead of one hard label per sample, letting a single synthetic sample encode information for multiple classes; reports 2-4% accuracy gains over hard-label distillation, over 96% MNIST accuracy from 10 distilled images and about 92% from just 5, and extends the same approach to distilling sequential text data, retaining near-original accuracy on IMDB sentiment analysis from 20 distilled sentences.