DD-Ranking: Rethinking the Evaluation of Dataset Distillation

Zekai Li, Xinhao Zhong, Samir Khaki, Zhiyuan Liang, Yuhao Zhou, Mingjia Shi, Ziqiao Wang, Xuanlei Zhao, Wangbo Zhao, Ziheng Qin, Mengxuan Wu, Pengfei Zhou, Haonan Wang, David Junhao Zhang, Jia-Wei Liu, Shaobo Wang, Dai Liu, Linfeng Zhang, Guang Li, Kun Wang, Zheng Zhu, Zhiheng Ma, Joey Tianyi Zhou, Jiancheng Lv, Yaochu Jin, Peihao Wang, Kaipeng Zhang, Lingjuan Lyu, Yiran Huang, Zeynep Akata, Zhiwei Deng, Xindi Wu, George Cazenavette, Yuzhang Shang, Justin Cui, Jindong Gu, Qian Zheng, Hao Ye, Shuo Wang, Xiaobo Wang, Yan Yan, Angela Yao, Mike Zheng Shou, Tianlong Chen, Hakan Bilen, Baharan Mirzasoleiman, Manolis Kellis, Konstantinos N. Plataniotis, Zhangyang Wang, Bo Zhao, Yang You, Kai Wang

arXiv 2025 · first public 2025-05-19 · arXiv 2505.13300

paper ↗code ↗project page ↗core✓ full text read

In one paragraph

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.

Explained

What came before

DC-BENCH (2022) standardized augmentation/architecture for hard-label matching methods (DC, DSA, DM, MTT), but by 2025 the field's frontier (SRe2L, RDED, CDA, DWA, D4M, EDC, G-VBSM, DATM) all train evaluation models with epoch-wise teacher soft labels and method-specific augmentation recipes (Table 1 catalogs wildly different epochs, batch sizes, optimizers and LR schedules across 13 methods), none of which DC-BENCH's protocol controls for.

The problem

Test accuracy conflates three things that DD papers report as one number: the informativeness of the synthetic images, the knowledge-distillation effect of soft labels from a pretrained teacher, and the effect of the post-hoc data augmentation/training recipe. The paper's diagnosis, backed by controlled experiments, is that random real images annotated with the same soft labels often beat the actual distilled images (Section 2.2), and a substantial share of reported gains vanishes when augmentation is removed (Table 2, especially at ImageNet-1K resolution).

The idea

Replace raw accuracy with two relative metrics that hold the confound fixed and vary only the data: a Label-Robust Score (LRS) that combines how much of the *hard-label* accuracy gap to the full dataset a method closes (Hard Label Recovery) with how much a method beats *random selection under its own label recipe* (Improvement over Random, IOR); and an Augmentation-Robust Score (ARS) that requires a method to beat random selection both with and without its augmentation.

How it works

No new synthesis method. Defines HLR = acc_real-hard - acc_syn-hard (smaller is better); IOR = acc_syn-any - acc_rdm-any, i.e. synthetic vs. a same-size random subset evaluated under the identical label type/recipe (larger is better, comparable across hard- and soft-label methods since both are relative to their own random-selection control); LRS = a normalized (via 100%*(e^alpha - e^-1)/(e - e^-1)) combination alpha = lambda*IOR - (1-lambda)*HLR, lambda=0.5 by default. ARS is defined analogously from IOR computed both with and without the method's augmentation (beta = gamma*IOR_aug + (1-gamma)*IOR_noaug). Benchmarks 13 methods (DC, DSA, MTT, DM, DataDAM as hard-label; SRe2L, DATM, EDF, DWA, RDED, CDA, EDC, G-VBSM as soft-label) on CIFAR-10, CIFAR-100, TinyImageNet and ImageNet-1K, each under its own reported architecture, then cross-checked on AlexNet/ResNet-50/VGG-11/Swin-Tiny/ViT-Base for robustness; 5 seeds per number.

Evidence

ImageNet-1K, soft-label/relabel+KD regime (Table 6, Section "Findings" i): current DD methods "consistently underperform random selection" (negative IOR) across most IPC settings once random images are given the same soft labels, and the gap widens as IPC grows. CIFAR-10, IPC1 (Table 3, lambda=0.5): SRe2L has IOR = -1.9 (worse than random under its own soft-label recipe) and RDED IOR = -0.5, while hard-label MTT has IOR = +16.7 and LRS 23.1, the best on CIFAR-10; at IPC50, SRe2L IOR = -11.5, RDED IOR = -1.5 — soft-label decoupled methods lose to random selection at every tested IPC on CIFAR-10 despite far higher raw accuracy. Augmentation-robust score, ImageNet-1K (Table 7): most decoupled/generative methods (SRe2L, D4M, DWA, CDA, G-VBSM) show negative IOR both with and without augmentation at IPC50 (e.g. SRe2L: -13.2 with aug, -21.0 without), and the gap to random widens as augmentation is removed and as IPC grows, meaning their accuracy edge is substantially manufactured by the augmentation recipe, not the images. Cross-architecture stability (Fig. 2) and cross-teacher stability (Fig. 3): LRS varies little across 4 evaluation architectures and across different teacher models used for relabeling, validating LRS as a more stable metric than raw accuracy. Among soft-label methods, DATM/EDF (soft labels *jointly optimized* with images during trajectory matching) beat D4M/SRe2L/RDED (labels generated post hoc by a frozen teacher) on LRS — matching-based methods remain the strongest baselines overall despite not scaling to ImageNet-1K.

Limitations

LRS and ARS still require access to the exact training recipe (optimizer, LR, epochs, augmentation) of each method to reproduce a matched random-selection control, so the benchmark is only as fair as each method's disclosed hyperparameters (Table 1 shows several methods use non-default optimizers without motivation). The paper reports ARS only for ImageNet-1K "due to space limit," leaving CIFAR/TinyImageNet augmentation-robustness less thoroughly documented. LRS's normalization (lambda=0.5 default) is a design choice; the appendix explores other lambda values (Tables 10-19) but the main paper does not argue lambda=0.5 is uniquely correct. Matching-based (hard-label) methods score best on LRS but the paper does not resolve whether this reflects genuinely more informative images or simply that hard-label evaluation has a smaller ceiling to close (full-dataset hard-label accuracy is lower, so HLR is easier to shrink). Does not evaluate coreset-selection methods (K-Center, forgetting) as first-class baselines with LRS/ARS, only as the "random" control.

Written by the atlas from the paper's full text. Check the paper for exact numbers.

Where it sits

Setting: Image classification

Design choices

Largest scale evaluated

mnist-cifar, tiny-imagenet, imagenet-1k

Abstract (verbatim from arXiv)

In recent years, dataset distillation has provided a reliable solution for data compression, where models trained on the resulting smaller synthetic datasets achieve performance comparable to those trained on the original datasets. To further improve the performance of synthetic datasets, various training pipelines and optimization objectives have been proposed, greatly advancing the field of dataset distillation. Recent decoupled dataset distillation methods introduce soft labels and stronger data augmentation during the post-evaluation phase and scale dataset distillation up to larger datasets (e.g., ImageNet-1K). However, this raises a question: Is accuracy still a reliable metric to fairly evaluate dataset distillation methods? Our empirical findings suggest that the performance improvements of these methods often stem from additional techniques rather than the inherent quality of the images themselves, with even randomly sampled images achieving superior results. Such misaligned evaluation settings severely hinder the development of DD. Therefore, we propose DD-Ranking, a unified evaluation framework, along with new general evaluation metrics to uncover the true performance improvements achieved by different methods. By refocusing on the actual information enhancement of distilled datasets, DD-Ranking provides a more comprehensive and fair evaluation standard for future research advancements.

BibTeX (generated; prefer the venue's official entry)

@article{li2025ranking,
  title   = {DD-Ranking: Rethinking the Evaluation of Dataset Distillation},
  author  = {Zekai Li and Xinhao Zhong and Samir Khaki and Zhiyuan Liang and Yuhao Zhou and Mingjia Shi and Ziqiao Wang and Xuanlei Zhao and Wangbo Zhao and Ziheng Qin and Mengxuan Wu and Pengfei Zhou and Haonan Wang and David Junhao Zhang and Jia-Wei Liu and Shaobo Wang and Dai Liu and Linfeng Zhang and Guang Li and Kun Wang and Zheng Zhu and Zhiheng Ma and Joey Tianyi Zhou and Jiancheng Lv and Yaochu Jin and Peihao Wang and Kaipeng Zhang and Lingjuan Lyu and Yiran Huang and Zeynep Akata and Zhiwei Deng and Xindi Wu and George Cazenavette and Yuzhang Shang and Justin Cui and Jindong Gu and Qian Zheng and Hao Ye and Shuo Wang and Xiaobo Wang and Yan Yan and Angela Yao and Mike Zheng Shou and Tianlong Chen and Hakan Bilen and Baharan Mirzasoleiman and Manolis Kellis and Konstantinos N. Plataniotis and Zhangyang Wang and Bo Zhao and Yang You and Kai Wang},
  journal = {arXiv preprint arXiv:2505.13300},
  year    = {2025}
}