Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Task-Specific Generative Dataset Distillation with Difficulty-Guided Sampling

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Sampling synthetic images to match the original dataset's difficulty distribution improves distilled-dataset classification accuracy.

desk verdict A plausible post-hoc sampling trick for generative dataset distillation, but the missing random-from-pool baseline currently confounds the central claim. read the letter →

arxiv 2507.03331 v2 pith:FIKT7FKC submitted 2025-07-04 cs.CV cs.AIcs.LG

Mingzhuo Li , Guang Li , Jiafeng Mao , Linfeng Ye , Takahiro Ogawa , Miki Haseyama This is my paper
classification cs.CVcs.AIcs.LG
keywords datasetdistillationgenerativedifficulty-guidedsamplingdiffusionmodelsimageclassificationdistributionInformationBottleneckNetsubsets
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Generative dataset distillation produces large pools of synthetic images, but existing methods usually choose the final compact set by matching the original data's overall distribution and ignore what the downstream task needs. This paper argues that for classification, the distilled set should reproduce the original dataset's distribution of sample difficulty, where difficulty means how unconfidently a pretrained classifier predicts each image's true label. To do this, it samples from a generated image pool so that the selected set's difficulty histogram resembles the original dataset's, and applies a logarithmic transformation to counteract the tendency of generative models to produce too many easy images. Across ImageWoof, ImageNette, and ImageIDC, the resulting sets train several architectures to higher accuracy than the underlying Minimax pipeline and other baselines at the same image-per-class budgets. The point is that task-specific information, namely how hard each sample is, can guide which synthetic images to keep, not just how to generate them.

What carries the argument

The machinery is a sampling step layered on top of the Minimax generative distillation pipeline. First, every image in the generated pool receives a difficulty score $D_x = 1 - P_{f_\theta}(y_{\mathrm{true}}|x)$ from a ResNet-50 trained on the full ImageNet dataset. Second, a sampling distribution is built by scaling the original dataset's difficulty histogram to the desired image-per-class count, so sampling from the pool reproduces the original proportion of easy and hard examples. Third, before that scaling, both histograms are processed by a logarithmic transform applied to a threshold-clipped version of the difficulty distribution, with thresholds selected by minimizing a KL-divergence trade-off between matching the original distribution and matching uniformity. The logarithmic correction is the load-bearing piece that makes the matching feasible, because raw generated pools are biased toward easy samples and would otherwise leave large parts of the difficulty range unrepresented.

What would settle it

Replace the ImageNet ResNet-50 difficulty scorer with a model trained from a different random seed or a different architecture, or temperature-scale its probabilities before computing $D_x$; if the downstream accuracy gains over the underlying Minimax pipeline vanish or reverse, the difficulty signal is an artifact of the scoring model rather than a property of the classification task.

Watch

Extended reading notes

Core claim

The central claim is that classification accuracy of generative dataset distillation improves when the final distilled set is selected according to the difficulty distribution of the original dataset, not just by the generic distributional alignment used during generation. The difficulty of an image $x$ is defined as $D_x = 1 - P_{f_\theta}(y_{\mathrm{true}}|x)$, the complement of the confidence a pretrained classifier assigns to the correct class. The sampling distribution is obtained by scaling the original dataset's difficulty histogram to the target number of images per class, so that the selected synthetic images contain the same proportions of easy and hard examples. Because generative models oversupply easy samples, the histogram is first modified by a logarithmic transformation whose clipping thresholds are chosen by a KL-divergence trade-off between uniformity and fidelity to the original distribution. As the paper reports it, this difficulty matching outperforms the Minimax baseline and other compared methods on ImageWoof, ImageNette, and ImageIDC at image-per-class values of 10, 20, and 50.

Load-bearing premise

The method's success rests on the assumption that difficulty scores from a single ImageNet-trained ResNet-50 reliably identify which synthetic images are hard for the downstream classifiers; if that scoring model is miscalibrated, memorizes the original images, or does not reflect the inductive biases of the test architectures, the matching target is misaligned and the reported gains may be artifacts of the scorer.

Editorial extensions

If this is right

  • At equal image-per-class budgets, difficulty-guided sampling improves classification accuracy over the underlying Minimax pipeline on the three ImageNet subsets; for example, on ImageWoof with ResNetAP-10 at IPC 50 accuracy rises from 54.4 to 57.1.
  • The scaling strategy that reproduces the original difficulty histogram beats four hand-shaped sampling distributions (hill, ground, slope, and cliff), which supports the paper's hypothesis that similarity to the original difficulty distribution is what drives the gain.
  • An image pool of five times the target set size gives the best accuracy among the tested pool sizes, so the method comes with a concrete practical recommendation for how much generation is needed.
  • Framed through the Information Bottleneck objective, the method explicitly improves $I(T;Y)$ rather than only $I(X;T)$, which the paper expects to extend to other downstream tasks once a suitable task-relevant proxy is defined.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit: the ImageNet ResNet-50 scorer could be replaced by an ensemble or by a model architecture-matched to the downstream test model; if gains persist, difficulty is a robust task signal, and if not, the scorer choice becomes a free parameter of the recipe.
  • The paper itself notes that the logarithmic thresholding distorts the original difficulty distribution, so a natural follow-up is to replace the KL-based threshold search with direct histogram transport, which could match the target distribution without the distortion the paper acknowledges.
  • Difficulty-guided sampling could be tested in continual or federated learning, where each client or task has its own difficulty profile; the paper's Markov-chain formulation suggests the same selection rule should transfer, but this is not demonstrated here.
  • One could test whether selecting for difficulty implicitly selects for shared visual prototypes: the paper's visualizations show that images of similar difficulty share common features, which would give a mechanistic explanation independent of histogram matching.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a task-specific sampling strategy for generative dataset distillation. Starting from an image pool of n × IPC images produced by the Minimax diffusion pipeline, it computes a per-image difficulty score D_x = 1 - P(y_true|x) using a pretrained ImageNet ResNet-50, applies a clipped logarithmic transformation to the difficulty distribution, and then selects IPC images so that the selected set matches the (transformed) difficulty distribution of the original dataset. The method is evaluated on three ImageNet subsets (ImageWoof, ImageNette, ImageIDC) with three downstream architectures at IPC 10, 20, and 50, and is claimed to outperform existing generative and non-generative distillation methods. The paper also includes ablations on the choice of sampling distribution (Table 3) and on the image-pool size multiplier n (Table 4).

Significance. If the central claim held, the paper would make a useful contribution to generative dataset distillation by showing that task-specific difficulty information can improve downstream classification over purely distribution-alignment-based generation. The idea is simple, the writing is clear, and the paper includes code, repeated experiments (3 seeds), and comparisons across three datasets and three architectures. However, the empirical evidence as presented does not yet isolate the effect of difficulty-guided sampling from the confounding effect of using a larger candidate pool, and several method hyperparameters are selected on the same benchmark that is used for the headline results. The Information Bottleneck framing and the log-transform correction are interesting but are not backed by a derivation that connects Eq. (5) to the actual sampling rule, so the contribution rests on the experimental comparison.

major comments (4)
  1. [§3.1, §3.3, Table 4] The reported comparison changes two variables at once. The Ours pipeline generates a 5×IPC image pool (Section 3.1 and Table 4) and then selects IPC images from it, whereas the Minimax baseline is the direct Minimax output at the target IPC. Thus Ours differs from Minimax both in the candidate-pool size and in the selection rule. A random-sampling-from-the-same-5×IPC-pool baseline is missing; without it, the improvement over Minimax could be caused by the larger pool providing better coverage of the generated distribution rather than by difficulty-guided selection. This control should be added to Tables 1 and 2, and ideally also to Table 3.
  2. [§3.3, §3.4, Eq. (9)] Several method choices are made post hoc on the same ImageWoof benchmark used for the headline results. The 'scale' distribution is chosen from Table 3, the pool-size multiplier n=5 is chosen from Table 4, and the clipping thresholds b,t and the weight λ in Eq. (9) are not reported at all. The paper should report the actual values used, state how they were selected, and provide evidence that the conclusions are not sensitive to these choices, for example by testing the selected configuration on a held-out subset or by reporting sensitivity curves for n and λ.
  3. [Tables 1 and 2] The accuracy differences are often within one standard deviation of the repeated runs, yet significance is claimed without any statistical test. For example, on ImageIDC at IPC=50, Minimax is 74.1±0.2 and Ours is 74.2±0.7, a negligible difference; on ImageWoof IPC=10 with ConvNet-6, Ours is 35.1±0.5 vs Minimax 34.1±0.4. With only three repetitions, the paper cannot support the claim that Ours systematically outperforms Minimax. Paired significance tests or additional seeds should be reported.
  4. [§2.2, §3.1, Eq. (6)] The difficulty oracle is a single ResNet-50 pretrained on full ImageNet, and the same oracle scores both the original images and the generated pool. The central assumption is that this oracle's confidence is a valid proxy for difficulty for all downstream architectures. The paper does not test whether results are sensitive to the choice of oracle, whether the oracle is calibrated on the generated images, or whether a different pretrained model would lead to the same sampling distribution. A robustness experiment with a different scoring model (e.g., ViT or a smaller ResNet) would strengthen the claim that the method targets task-relevant information rather than artifacts of one particular classifier.
minor comments (6)
  1. [Eq. (2)] The training objective in Eq. (2) is written as arg max over θ, but the noise-prediction loss is minimized; it should be arg min.
  2. [§2.3, Eq. (7)] The notation P_X(n) is ambiguous: n is used both as the pool-size multiplier and as the index of the difficulty distribution, and N is not clearly defined as the number of bins. This should be clarified.
  3. [References] Reference [32] and [33] are the same paper (Rombach et al., latent diffusion) with different titles; one should be removed or corrected.
  4. [Author affiliation] The affiliation contains a typo: 'Hokkaido Univerisy' should be 'Hokkaido University'.
  5. [§3.3] The statement that 'smaller sampled datasets with a higher proportion of easier samples, as well as larger sampled datasets with a higher proportion of more difficult samples, tend to yield better performance' is hard to interpret without a quantitative definition of the easy/difficult proportion; Figure 4 is only a sketch.
  6. [§2.2, Eq. (5)] The Information Bottleneck formulation is used as motivation, but Eq. (5) is never connected to the difficulty-based sampling rule; the paper should either make this link explicit or describe the IB discussion as a heuristic.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: difficulty matching is a design hypothesis, and reported accuracy is measured on held-out data rather than defined by the match.

full rationale

The paper's central claim is that sampling a generated image pool to match the difficulty distribution of the original dataset improves downstream classification accuracy. This is presented explicitly as a hypothesis in Section 2.2: 'we hypothesize that a distilled dataset exhibiting a similar difficulty distribution to the original one may yield improved performance.' The difficulty target D_x = 1 - P(y_true|x) is a statistic of the original dataset computed with a pretrained ImageNet ResNet-50, and the final dataset is selected from a Minimax-generated pool by scaling that distribution. However, the reported accuracy in Tables 1 and 2 is measured on held-out validation splits, which is not a function of the difficulty-matching objective. Therefore, no prediction reduces by construction to a fitted value. The logarithmic transformation and threshold selection in Section 2.3 are heuristic preprocessing steps, and the manuscript itself acknowledges in Section 3.2 that the transformation 'alters the original dataset's difficulty distribution after transformation, highlighting the need for further discussion.' The citation to Wang et al. [40] includes a co-author but is used as motivation and as an observation of easy-sample bias; Fig. 2 provides in-paper evidence of the bias, so the self-citation is not load-bearing for the main empirical claim. The main validity concerns are missing a random-sampling-from-pool control and selecting n=5 and the 'scale' distribution on the same ImageWoof benchmark where headline results are reported. Those are experimental design issues, not circular derivation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new entities are introduced; difficulty is a scalar derived from an existing pretrained model. The main costs are the free parameters and domain assumptions listed above, especially the choice of the difficulty oracle and the target distribution.

free parameters (3)
  • lambda in Eq. 9 = not reported
    Weights the KL divergence toward the original distribution versus the uniform target; no value or sensitivity analysis is given.
  • clipping thresholds b,t = optimized via Eq. 9
    Chosen to balance KL to original and uniform; these clip the difficulty histograms and directly change the sampling distribution.
  • image pool size multiplier n = 5
    Selected from Table 4 based on higher-IPC results; affects diversity and redundancy of the pool and the thresholds.
assumptions (5)
  • domain assumption Information Bottleneck formulation applies to dataset distillation with Markov chain X -> T -> Y.
    Invoked in Section 2.2 to justify maximizing I(T;Y), but the method does not optimize IB directly.
  • domain assumption Inverse confidence from a pretrained ResNet-50 is a valid measure of classification difficulty.
    Eq. 6 defines difficulty this way; no calibration or robustness check is provided.
  • ad hoc to paper Matching the original dataset's difficulty distribution improves downstream accuracy.
    Stated as hypothesis in Section 2.2 and supported only by internal comparisons in Table 3.
  • ad hoc to paper Logarithmic transformation toward a uniform distribution corrects generator bias without destroying task-relevant structure.
    Section 2.3 assumes the easy-sample bias is best corrected by flattening; the paper acknowledges this alters the original distribution.
  • domain assumption The generated image pool covers all difficulty bins after transformation.
    Required for matching to be feasible; Fig. 2 suggests coverage is imperfect and some intervals remain unrepresented.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Specific Generative Dataset Distillation with Difficulty-Guided Sampling." pith.science (2026). https://pith.science/paper/FIKT7FKC

@misc{pith2026250703331,
  author       = {Pith},
  title        = {Pith review of: Task-Specific Generative Dataset Distillation with Difficulty-Guided Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FIKT7FKC}},
  note         = {Machine review of arXiv:2507.03331}
}
read the original abstract

To alleviate the reliance of deep neural networks on large-scale datasets, dataset distillation aims to generate compact, high-quality synthetic datasets that can achieve comparable performance to the original dataset. The integration of generative models has significantly advanced this field. However, existing approaches primarily focus on aligning the distilled dataset with the original one, often overlooking task-specific information that can be critical for optimal downstream performance. In this paper, focusing on the downstream task of classification, we propose a task-specific sampling strategy for generative dataset distillation that incorporates the concept of difficulty to consider the requirements of the target task better. The final dataset is sampled from a larger image pool with a sampling distribution obtained by matching the difficulty distribution of the original dataset. A logarithmic transformation is applied as a pre-processing step to correct for distributional bias. The results of extensive experiments demonstrate the effectiveness of our method and suggest its potential for enhancing performance on other downstream tasks. The code is available at https://github.com/SumomoTaku/DiffGuideSamp.

Figures

Figures reproduced from arXiv: 2507.03331 by the authors.

Figure 1
Figure 1. The workflow of the proposed method with the overall linear process marked as green. The generation of the image pool is [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The difficulty distributions of different datasets, in the example of two classes of ImageWoof. The x-axis represents difficulty [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of images of the original and distilled dataset with difficulty scores. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of pre-defined sampling distributions un [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 38 canonical work pages

  1. [1]

    Efros, and Jun-Yan Zhu

    George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. InProc. CVPR, pages 10718– 10727, 2022. 1

    work page 2022

  2. [2]

    Efros, and Jun-Yan Zhu

    George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Generalizing dataset dis- tillation via deep generative prior. InProc. CVPR, pages 3739–3748, 2023. 1

    work page 2023

  3. [3]

    Chan-Santiago, Praveen Tirupattur, Gaurav Ku- mar Nayak, Gaowen Liu, and Mubarak Shah

    Jeffrey A. Chan-Santiago, Praveen Tirupattur, Gaurav Ku- mar Nayak, Gaowen Liu, and Mubarak Shah. MGD3: Mode- guided dataset distillation using diffusion models. InProc. ICML, 2025. 3

    work page 2025

  4. [4]

    Provable and efficient dataset distillation for kernel ridge regression

    Yilan Chen, Wei Huang, and Tsui-Wei Weng. Provable and efficient dataset distillation for kernel ridge regression. In Proc. NeurIPS, 2024. 1

    work page 2024

  5. [5]

    OPTICAL: Leveraging optimal transport for contribution allocation in dataset distillation

    Xiao Cui, Yulei Qin, Wengang Zhou, Hongsheng Li, and Houqiang Li. OPTICAL: Leveraging optimal transport for contribution allocation in dataset distillation. InProc. CVPR,

    work page

  6. [6]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. InProc. CVPR, pages 248–255, 2009. 5

    work page 2009

  7. [7]

    and Welling Max

    Kingma Diederik P. and Welling Max. Auto-encoding vari- ational bayes.arXiv preprint arXiv:1312.6114, pages 1–14,

    arXiv

  8. [8]

    Calibrated dataset condensation for faster hyperparameter search.arXiv preprint arXiv:2405.17535,

    Mucong Ding, Yuancheng Xu, Tahseen Rabbani, Xiaoyu Liu, Brian Gravelle, Teresa Ranadive, Tai-Ching Tuan, and Furong Huang. Calibrated dataset condensation for faster hyperparameter search.arXiv preprint arXiv:2405.17535,

    arXiv

Show all 46 references
  1. [9]

    imagenette.https://github.com/fastai/ imagenette, 2019

    Fastai. imagenette.https://github.com/fastai/ imagenette, 2019. 5

    2019

  2. [10]

    Efficient deep learning: A survey on mak- ing deep learning models smaller, faster, and better.ACM Computing Surveys, 55(12):1–37, 2023

    Menghani Gaurav. Efficient deep learning: A survey on mak- ing deep learning models smaller, faster, and better.ACM Computing Surveys, 55(12):1–37, 2023. 1

    2023

  3. [11]

    Dynamic few-shot visual learning without forgetting

    Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. InProc. CVPR, pages 4367–4375, 2018. 5

    2018

  4. [12]

    Efficient dataset distillation via minimax diffusion

    Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Hao- nan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. InProc. CVPR, pages 15793–15803, 2024. 1, 2, 5

    2024

  5. [13]

    Sum- marizing stream data for memory-restricted online continual learning

    Jianyang Gu, Kai Wang, Wei Jiang, and Yang You. Sum- marizing stream data for memory-restricted online continual learning. InProc. AAAI, 2024. 1

    2024

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProc. CVPR, pages 770–778, 2016. 5

    2016

  7. [15]

    Unlocking the potential of federated learning: The sym- phony of dataset distillation via deep generative latents

    Yuqi Jia, Saeed Vahidian, Jingwei Sun, Jianyi Zhang, Vyach- eslav Kungurtsev, Neil Zhenqiang Gong, and Yiran Chen. Unlocking the potential of federated learning: The sym- phony of dataset distillation via deep generative latents. In Proc. ECCV, 2024. 1

    2024

  8. [16]

    Dataset condensation via efficient synthetic- data parameterization

    Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic- data parameterization. InProc. ICML, pages 11102–11118,

  9. [17]

    Soft-label anonymous gastric x-ray image distillation

    Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Soft-label anonymous gastric x-ray image distillation. In Proc. ICIP, pages 305–309, 2020. 1

    2020

  10. [18]

    Compressed gastric image generation based on soft-label dataset distillation for medical data sharing.Computer Meth- ods and Programs in Biomedicine, 227:107189, 2022

    Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Compressed gastric image generation based on soft-label dataset distillation for medical data sharing.Computer Meth- ods and Programs in Biomedicine, 227:107189, 2022. 1

    2022

  11. [19]

    Awesome dataset distillation.https : / / github

    Guang Li, Bo Zhao, and Tongzhou Wang. Awesome dataset distillation.https : / / github . com / Guang000 / Awesome-Dataset-Distillation, 2022. 1

    2022

  12. [20]

    Dataset distillation using parameter pruning.IEICE Transac- tions on Fundamentals of Electronics, Communications and Computer Sciences, 2023

    Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Dataset distillation using parameter pruning.IEICE Transac- tions on Fundamentals of Electronics, Communications and Computer Sciences, 2023. 1

    2023

  13. [21]

    Dataset distillation for medical dataset sharing

    Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Dataset distillation for medical dataset sharing. InProc. AAAI Workshop, pages 1–6, 2023. 1

    2023

  14. [22]

    Importance-aware adaptive dataset distillation.Neural Net- works, 2024

    Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama. Importance-aware adaptive dataset distillation.Neural Net- works, 2024. 1

    2024

  15. [23]

    Generative dataset distillation: Balancing global structure and local details

    Longzhen Li, Guang Li, Ren Togo, Keisuke Maeda, Takahiro Ogawa, and Miki Haseyama. Generative dataset distillation: Balancing global structure and local details. In Proc. CVPR Workshop, pages 7664–7671, 2024. 1

    2024

  16. [24]

    Generative dataset distillation based on self-knowledge distillation

    Longzhen Li, Guang Li, Ren Togo, Keisuke Maeda, Takahiro Ogawa, and Miki Haseyama. Generative dataset distillation based on self-knowledge distillation. InProc. ICASSP, 2025. 1

    2025

  17. [25]

    Diversity-driven generative dataset distilla- tion based on diffusion model with self-adaptive memory

    Mingzhuo Li, Guang Li, Jiafeng Mao, Takahiro Ogawa, and Miki Haseyama. Diversity-driven generative dataset distilla- tion based on diffusion model with self-adaptive memory. In Proc. ICIP, 2025. 1

    2025

  18. [26]

    Hyperbolic dataset distillation.arXiv preprint arXiv:2505.24623, 2025

    Wenyuan Li, Guang Li, Keisuke Maeda, Takahiro Ogawa, and Miki Haseyama. Hyperbolic dataset distillation.arXiv preprint arXiv:2505.24623, 2025. 1

    2025

  19. [27]

    The evolution of dataset distillation: Toward scalable and generalizable solutions.arXiv preprint arXiv:2502.05673, 2025

    Ping Liu and Jiawei Du. The evolution of dataset distillation: Toward scalable and generalizable solutions.arXiv preprint arXiv:2502.05673, 2025. 1

    2025 arXiv

  20. [28]

    Reducing catas- trophic forgetting with learning on synthetic data

    Wojciech Masarczyk and Ivona Tautkute. Reducing catas- trophic forgetting with learning on synthetic data. InProc. CVPR Workshop, pages 4321–4326, 2020. 1

    2020

  21. [29]

    Understanding of machine learn- ing with deep learning: architectures, workflow, applications and future directions.Computers, 12(5):1–27, 2023

    Taye Mohammad Mustafa. Understanding of machine learn- ing with deep learning: architectures, workflow, applications and future directions.Computers, 12(5):1–27, 2023. 1

    2023

  22. [30]

    Dataset meta-learning from kernel ridge-regression

    Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. InProc. ICLR,

  23. [31]

    Scalable diffusion models with transformers.arXiv preprint arXiv:2212.09748, pages 1–25, 2023

    William Peebles and Saining Xie. Scalable diffusion models with transformers.arXiv preprint arXiv:2212.09748, pages 1–25, 2023. 5

    2023 arXiv

  24. [32]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn. Ommer. High-resolution image synthesis with latent diffusion models. InProc. CVPR, pages 10684–10695, 2022. 2

    2022

  25. [33]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn. Ommer. High-resolution image synthesis with latent diffusion models. InProc. CVPR, pages 10684–10695, 2022. 1

    2022

  26. [34]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Sali- mans, David J. Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713– 4726, 2023. 1

    2023

  27. [35]

    Liu, Yuri A

    Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z. Liu, Yuri A. Lawryshyn, and Konstantinos N. Platanio- tis. DataDAM: Efficient dataset distillation with attention matching. InProc. ICCV, pages 17097–17107, 2023. 1

    2023

  28. [36]

    Active learning for convolu- tional neural networks: A core-set approach.arXiv preprint arXiv:1708.00489, pages 1–13, 2017

    Ozan Sener and Silvio Savarese. Active learning for convolu- tional neural networks: A core-set approach.arXiv preprint arXiv:1708.00489, pages 1–13, 2017. 5

    2017 arXiv

  29. [37]

    Generative dataset distillation based on diffusion model

    Duo Su, Junjie Hou, Guang Li, Ren Togo, Rui Song, Takahiro Ogawa, and Miki Haseyama. Generative dataset distillation based on diffusion model. InProc. ECCV Work- shop, 2024. 1

    2024

  30. [38]

    Generative dataset distillation based on diffusion model

    Duo Su, Junjie Hou, Guang Li, Ren Togo, Rui Song, Takahiro Ogawa, and Miki Haseyama. Generative dataset distillation based on diffusion model. InProc. ECCV Work- shop, pages 1–12, 2024. 1

    2024

  31. [39]

    Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset distillation.arXiv preprint arXiv:1811.10959, pages 1–14, 2018. 1

    2018 arXiv

  32. [40]

    Training data synthesis with difficulty controlled dif- fusion model.arXiv preprint arXiv:2411.18109, pages 1–10,

    Zerun Wang, Jiafeng Mao, Xueting Wang, and Toshihiko Ya- masaki. Training data synthesis with difficulty controlled dif- fusion model.arXiv preprint arXiv:2411.18109, pages 1–10,

  33. [41]

    Herding dynamical weights to learn

    Max Welling. Herding dynamical weights to learn. InProc. ICML, pages 1121–1128, 2009. 5

    2009

  34. [42]

    Difffit: Un- locking transferability of large diffusion models via simple parameter-efficient fine-tuning

    Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Un- locking transferability of large diffusion models via simple parameter-efficient fine-tuning. InProc. ICCV, pages 4207– 4216, 2023. 5

    2023

  35. [43]

    Dataset condensation with gradi- ent matching

    Bo Zhao and Hakan Bilen. Dataset condensation with gradi- ent matching. InProc. ICLR, pages 1–20, 2021. 1

    2021

  36. [44]

    Synthesizing informative training samples with gan

    Bo Zhao and Hakan Bilen. Synthesizing informative training samples with gan. InProc. NeurIPS Workshop, 2022. 1

    2022

  37. [45]

    Dataset condensation with distri- bution matching

    Bo Zhao and Hakan Bilen. Dataset condensation with distri- bution matching. InProc. WACV, pages 6514–6523, 2023. 1, 5

    2023

  38. [46]

    Differentially private dataset condensation

    Tianhang Zheng and Baochun Li. Differentially private dataset condensation. InProc. NDSS Workshop, 2024. 1

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.