Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that replacing oldest-in-first-out memory updates with eviction of the most cosine-similar latent makes diffusion-based generative dataset distillation produce more diverse, more representative distilled datasets…

desk verdict Useful incremental memory-update trick for diffusion-based dataset distillation, but the key ablation against Minimax's FIFO rule is missing, so the central claim is not yet demonstrated. read the letter →

arxiv 2505.19469 v1 pith:SF5QLHGF submitted 2025-05-26 cs.LG cs.AIcs.CV

classification cs.LGcs.AIcs.CV
keywords datasetdistillationdiffusionmodelself-adaptivememorydiversitygenerativelatentImageWoofdownstreamvalidationaccuracy
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

Dataset distillation compresses a large training set into a small synthetic set that can still train a model to comparable accuracy. Generative distillation methods that train a diffusion model to produce the synthetic set can fail to cover the original data's spread, because the generated images cluster into a few regions. This paper argues that the auxiliary memory guiding such distillation should be updated by removing the most redundant latent rather than the oldest one. With that self-adaptive memory, the distilled dataset becomes more diverse and more representative, and the paper reports higher downstream validation accuracy than prior methods on ImageWoof, ImageNette, and ImageIDC across several images-per-class settings.

What carries the argument

The self-adaptive memory is the central mechanism: two bounded buffers, one of real latents and one of generated latents, maintained during the eight-epoch distillation run. When a buffer exceeds its size, the latent whose summed cosine similarity to every other element is largest is evicted, so the buffer keeps covering the distribution instead of drifting toward a random subset. The buffers feed the two alignment losses: $L_{\mathrm{real}}$ pulls each generated latent toward its least similar stored real latent, and $L_{\mathrm{gen}}$ pushes it away from its most similar stored generated latent; together with the diffusion loss $L_{\mathrm{diffusion}}$ these form the objective $L = L_{\mathrm{diffusion}} + \lambda_r L_{\mathrm{real}} + \lambda_g L_{\mathrm{gen}}$ (Eq. 6). The cosine similarity vector is the diversity proxy that drives both the losses and the eviction decisions.

What would settle it

A direct falsifier would be to replace the cosine-similarity proxy with an independent diversity measure, such as clustering-based coverage over held-out real latents, and test whether the max-similarity eviction still improves accuracy: if the gain disappears under a different proxy, the paper's diversity explanation is wrong. A simpler falsifier is to keep the same losses and memory size but evict a random latent instead of the most-similar one; if random eviction matches the reported accuracy, then the eviction rule itself is not what drives the improvement.

Watch

Extended reading notes

Core claim

The paper's claim is that diversity in diffusion-based generative dataset distillation can be controlled directly by the memory that feeds two similarity losses. In the real-memory loss the method maximizes the minimum cosine similarity between a generated latent and the stored real latents (Eq. 4), while in the generative-memory loss it minimizes the maximum cosine similarity between the generated latent and stored generated latents (Eq. 5). The new ingredient is the eviction rule: when a memory exceeds its fixed size, it computes a similarity vector by summing each element's cosine similarity to all other elements and removes the element with the largest sum (Eqs. 7 and 8), instead of popping the oldest latent. The paper argues this keeps both memories spread across the original distribution, and reports that on ImageWoof this raises accuracy from 51.3 to 52.7 at IPC 50 with ConvNet-6 and from 55.1 to 56.9 with ResNetAP-10, with superior accuracy across all tested settings on ImageWoof and most settings on the other subsets.

Load-bearing premise

The load-bearing premise is that cosine similarity between latents in the VAE/DiT feature space faithfully measures whether the distilled dataset represents the original distribution for downstream classification, and that evicting the most-similar latent therefore improves representativeness; if this proxy misleads, the memory updates and the two alignment losses can push generation in ways that do not help, or even hurt, validation accuracy.

Editorial extensions

If this is right

  • For a fixed distillation budget, downstream accuracy can be improved simply by changing how the auxiliary memory is maintained, without altering the diffusion model or the number of generated images.
  • The largest gains appear at low images-per-class settings, where a distilled set is most likely to miss parts of the original distribution, so the method addresses the regime where diversity matters most.
  • Because the memory stores latents and keeps a fixed size of 64, the extra cost is small, which makes the approach practical for repeated distillation in continual learning, federated learning, and privacy-preserving data sharing.
  • On the easy ImageNette classes the gain over the baseline shrinks or reverses at high IPC, which is consistent with the claim that diversity pressure matters only when the original classes are hard to separate.
  • The same improvement pattern across three models (ConvNet-6, ResNetAP-10, ResNet-18) suggests the diversity gain is not tied to one architecture.

Reading between the lines

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

  • A testable extension the paper does not run: measure the memory's coverage directly, for example by clustering the stored latents or by comparing them with held-out real latents, and check whether coverage and downstream accuracy move together under different eviction policies.
  • If cosine similarity in VAE/DiT space is the right diversity proxy, the same eviction rule could transfer to other generative distillation frameworks that maintain a memory, including non-latent diffusion or GAN-based methods, without retraining the core model.
  • The ablation shows that generative-memory eviction alone accounts for most of the gain, suggesting that the real-memory policy helps mainly at low IPC; an implicit prediction is that at high IPC the real-memory policy could be relaxed or removed with little loss, which the paper does not test.
  • The paper's insensitivity to memory size predicts that even much smaller memories, such as 16 latents, should retain most of the improvement; a direct check would separate the value of the eviction rule from the value of sheer memory capacity.
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

3 major / 5 minor

Summary. This paper proposes a diversity-driven generative dataset distillation method. Building on the Minimax diffusion-based distillation framework, the authors introduce a self-adaptive memory consisting of real and generated latents. Instead of evicting the oldest element as in Minimax, the proposed update rule removes the latent with the maximum total cosine similarity to all other elements in the memory (Eqs. 7-8). The memory is used to compute two additional loss terms, L_real and L_gen, which respectively pull the generated latents toward the least similar real latent and push them away from the most similar generated latent. The combined objective (Eq. 6) is used to fine-tune a DiT model. Experiments on ImageWoof, ImageNette, and ImageIDC report downstream validation accuracy improvements over Minimax and other baselines across several IPC settings. An ablation in Table 3 compares min/max variants of the similarity-based eviction rule, and a hyperparameter analysis in Fig. 4 examines λ_r, λ_g, and memory size.

Significance. If the proposed mechanism is validated, the self-adaptive memory rule would be a simple and effective modification to Minimax, offering consistent gains on three ImageNet subsets with modest computational overhead. The paper includes useful elements: three datasets, three test architectures, standard deviations from three runs, a strategy ablation, and a hyperparameter sensitivity study. However, the significance is currently limited because the central claim—that the max-similarity eviction rule improves diversity and hence accuracy—is not directly tested. The gains over Minimax could stem from other pipeline differences, and the 'diversity' improvement is inferred only from downstream accuracy, not measured. The cosine similarity proxy is also unvalidated.

major comments (3)
  1. [Sec. 3.3, Table 3] The ablation does not include a FIFO (oldest-out) baseline, which is the update rule used by Minimax. The four rows only vary the min/max selection for the real and generated memories; all are similarity-based policies. Since the paper's central contribution is the replacement of FIFO with max-similarity eviction, the comparison against Minimax in Tables 1-2 is not a controlled experiment. To claim that the self-adaptive memory causes the accuracy gains, the authors must add a FIFO condition within the same training pipeline (same DiT setup, epochs, batch size) and show that max-similarity eviction outperforms it.
  2. [Secs. 2.2-2.3, Eqs. (3)-(8)] The paper never measures diversity directly. The mechanism is that the cosine-similarity-based losses and eviction rule increase the diversity of the distilled dataset, but the only quantitative evidence is downstream validation accuracy. No diversity metric (e.g., average pairwise latent distance, coverage of the real distribution, or number of effective modes) is reported. The qualitative visualization in Fig. 3 is not sufficient. The authors should report a diversity metric on the distilled datasets produced by Minimax and by the proposed method to substantiate the causal link.
  3. [Sec. 3.4, Fig. 4] The hyperparameters λ_r, λ_g, and memory size are chosen using sweeps on ImageWoof, which is also the primary benchmark in Table 1. Even though the parameters are then applied to other datasets, the ImageWoof results are thus not independent of model selection. The authors should clarify whether any held-out validation was used, or at least acknowledge that the ImageWoof numbers reflect post-hoc selection. This practice can inflate accuracy estimates and is a particular concern because the reported improvements over Minimax are within 1-2 percentage points.
minor comments (5)
  1. [Eq. (3)] Eq. (3) is written as an arg max objective that is not used; the notation is misleading because it defines L_diversity as an operator rather than a loss. The authors should rephrase or remove this equation.
  2. [Algorithm 1, lines 7-15] The new latent is first enqueued and then the element with the highest total similarity is popped. If the newly added latent is itself the most similar, it is immediately removed, which may reduce the effective memory update. The authors should discuss or justify this design choice.
  3. [Table 2, Sec. 3.2] On ImageNette at IPC=50, Random achieves 76.1±1.1 versus the proposed method's 75.5±0.3; the paper's statement that the method 'demonstrates superior accuracy' should be qualified in the text.
  4. [General] The paper does not provide a public code release; given the sensitivity to training details, code would substantially improve reproducibility.
  5. [References] The references are largely appropriate, but the paper should cite the original Minimax paper more precisely and state which hyperparameters from Minimax are adopted.

Circularity Check

1 steps flagged · score 2.0 of 10

No central circularity: the headline accuracy gains are measured on external benchmarks, but the 'diversity' claim is partly definitional because diversity is defined as the cosine similarity that the objectives and memory update themselves optimize.

  1. self definitional [Sec. 2.2 Eq. (3); Sec. 2.3 Eqs. (7)-(8)]
    "We simplify the diversity as the similarity between the distilled and original distribution... For cosine similarity, higher scores indicate greater similarity between elements. Therefore, latents with higher similarity degrees, which may overly cluster the distribution when integrated, are prioritized for removal to maintain diversity."

    Diversity is defined in Eq. (3) as cosine similarity between latents, and the same cosine similarity is used in the optimization objectives (Eqs. (4)-(5)) and in the memory eviction rule (Eqs. (7)-(8)). The eviction rule removes exactly the latent with the largest summed cosine similarity, i.e., the element that most reduces the paper's own diversity measure. Thus the statement that the self-adaptive memory 'maintains diversity' is true by construction rather than by independent empirical demonstration; no separate diversity metric is reported. This does not force the downstream accuracy numbers, which are measured externally, so the circularity is limited to the diversity sub-claim.

full rationale

The paper's central numerical claims are downstream validation accuracies on ImageWoof, ImageNette, and ImageIDC, which are external benchmarks; no loss term equals the reported accuracy, and the self-citations (refs. [10], [11], [16], [19], [20], [24]) appear only as related work and are not load-bearing. The one definitional step is the treatment of 'diversity' itself: Eq. (3) equates diversity with cosine similarity, and the same metric drives both the losses and the self-adaptive eviction rule, so the claim of improved diversity is largely a restatement of the algorithm rather than an independent finding. The lack of a FIFO (oldest-out) control in Table 3 and hyperparameter selection on the evaluation benchmarks are experimental limitations that could affect attribution of the accuracy gains, but they are not circularity: the accuracy comparison itself is self-contained against external validation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central method depends on three unproven modeling choices: latent cosine similarity as a diversity metric, a 64-slot memory as a sufficient surrogate for the full dataset, and a greedy eviction rule that preserves representativeness. The diffusion loss and VAE/DiT are standard building blocks taken from prior work. The free parameters are the two loss weights and the memory size, selected via the hyperparameter analysis in Fig. 4 on the same evaluation benchmark.

free parameters (3)
  • lambda_r = 0.002
    Weight for real-memory alignment loss L_real; selected from the sweep in Fig. 4(a) on ImageWoof/ResNetAP-10, the same setting used in primary results.
  • lambda_g = 0.008
    Weight for generative repulsion loss L_gen; selected from Fig. 4(b).
  • memory_size (NR and NG) = 64
    Memory capacity for both real and generated latents; chosen from Fig. 4(c), where accuracy is reported to be insensitive across sizes.
assumptions (4)
  • domain assumption Cosine similarity in VAE/DiT latent space measures distributional diversity and representativeness for downstream classification.
    Used in Eqs. 4-5 to define L_real and L_gen; no validation that this proxy tracks classification-relevant diversity.
  • ad hoc to paper A 64-element memory of real latents, after max-similarity eviction, faithfully represents the full original dataset distribution.
    Sec. 2.3 states memory is used because the original dataset is too large; no analysis of approximation error or dependence on memory size beyond Fig. 4c.
  • domain assumption The pre-trained DiT and VAE encoders are fixed and their latent geometry is semantically meaningful for the target datasets.
    Sec. 2.1 adopts DiT and VAE without fine-tuning the encoder; the method relies on transferable latent features.
  • domain assumption Gradients of the min/max cosine-similarity objectives can be back-propagated through the diffusion model's generation path to update theta.
    Eq. 6 combines L_real and L_gen with L_diffusion; the paper never specifies the gradient path through the sampling or denoising process, which is nontrivial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory." pith.science (2026). https://pith.science/paper/SF5QLHGF

@misc{pith2026250519469,
  author       = {Pith},
  title        = {Pith review of: Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SF5QLHGF}},
  note         = {Machine review of arXiv:2505.19469}
}
read the original abstract

Dataset distillation enables the training of deep neural networks with comparable performance in significantly reduced time by compressing large datasets into small and representative ones. Although the introduction of generative models has made great achievements in this field, the distributions of their distilled datasets are not diverse enough to represent the original ones, leading to a decrease in downstream validation accuracy. In this paper, we present a diversity-driven generative dataset distillation method based on a diffusion model to solve this problem. We introduce self-adaptive memory to align the distribution between distilled and real datasets, assessing the representativeness. The degree of alignment leads the diffusion model to generate more diverse datasets during the distillation process. Extensive experiments show that our method outperforms existing state-of-the-art methods in most situations, proving its ability to tackle dataset distillation tasks.

Figures

Figures reproduced from arXiv: 2505.19469 by the authors.

Figure 2
Figure 2. The distillation process of our method. Randomly selected images are input to the diffusion model to obtain the generated latents and then the diffusion loss Ldiffusion. Two memory sets consisting of real and generated latents, respectively, assist in calculating diversity loss Lreal and Lgen. The memory takes self-adaptive updates based on the similarity vector after each epoch. memory is introduced to substitute t… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 32 canonical work pages

  1. [1]

    Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory

    INTRODUCTION The rapid development of deep learning leads to the widespread pop- ularity and adoption of deep neural networks, which make full use of large amounts of data and achieve remarkable results in the field of computer vision [1]. However, the reliance on data extends the training process to many hours or even days and imposes substantial demands...

  2. [2]

    Our method mainly includes three parts: preliminary diffusion model, optimization objectives for dataset distillation, and self-adaptive memory

    DIVERSITY-DRIVEN DA TASET DISTILLA TION The proposed method aims to obtain distilled datasets that better rep- resent the original ones by using diversity-driven optimization with the assistance of auxiliary memory. Our method mainly includes three parts: preliminary diffusion model, optimization objectives for dataset distillation, and self-adaptive memo...

  3. [3]

    EXPERIMENTS 3.1. Datasets and Evaluation Extensive experiments have been conducted to verify the effective- ness of the proposed method on three benchmark subsets of the full-sized ImageNet [30] dataset: ImageWoof [31], ImageNette [31], and ImageIDC [8]. ImageWoof contains 10 classes of specific dog Table 1. Comparison of downstream validation accuracy wi...

  4. [4]

    CONCLUSION In this paper, we have proposed a diversity-driven generative dataset distillation method based on a diffusion model. A key innovation of our method is the introduction of two self-adaptive memory sets for evaluating and improving the diversity of the distilled dataset, aiming to enhance its ability to represent the original dataset. Self- adap...

  5. [5]

    A survey on data-efficient algorithms in big data era,

    Amina Adadi, “A survey on data-efficient algorithms in big data era,”Journal of Big Data, vol. 8, no. 1, pp. 1–24, 2021

  6. [6]

    Review of deep learning: Concepts, cnn architectures, challenges, applications, future directions,

    Laith Alzubaidi, Jinglan Zhang, and et al., “Review of deep learning: Concepts, cnn architectures, challenges, applications, future directions,”Journal of Big Data, vol. 8, pp. 1–74, 2021

  7. [7]

    Dataset distillation,

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

  8. [8]

    Scalable diffusion models with transformers,

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

Show all 38 references
  1. [9]

    Awesome dataset distillation,

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

  2. [10]

    A compre- hensive survey to dataset distillation,

    Ruonan Yu, Songhua Liu, and Xinchao Wang, “A compre- hensive survey to dataset distillation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 1, pp. 150–170, 2023

  3. [11]

    Dataset condensation with gradi- ent matching,

    Bo Zhao and Hakan Bilen, “Dataset condensation with gradi- ent matching,” inProc. ICLR, 2021, pp. 1–20

  4. [12]

    Dataset condensation via efficient synthetic-data pa- rameterization,

    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 pa- rameterization,” inProc. ICML, 2022, pp. 11102–11118

  5. [13]

    Dataset distillation by matching training trajectories,

    George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu, “Dataset distillation by matching training trajectories,” inProc. CVPR, 2022, pp. 10718–10727

  6. [14]

    Dataset distillation using parameter pruning,

    Guang Li, Ren Togo, Takahiro Ogawa, and Miki Haseyama, “Dataset distillation using parameter pruning,”IEICE Trans- actions on Fundamentals of Electronics, Communications and Computer Sciences, 2023

  7. [15]

    Importance-aware adaptive dataset distillation,

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

  8. [16]

    Dataset meta-learning from kernel ridge-regression,

    Timothy Nguyen, Zhourong Chen, and Jaehoon Lee, “Dataset meta-learning from kernel ridge-regression,” inProc. ICLR, 2021, pp. 1–25

  9. [17]

    Dataset distillation with infinitely wide convolutional networks,

    Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee, “Dataset distillation with infinitely wide convolutional networks,” inProc. NeurIPS, 2021, pp. 5186–5198

  10. [18]

    Synthesizing informative training samples with gan,

    Bo Zhao and Hakan Bilen, “Synthesizing informative training samples with gan,” inProc. NeurIPS Workshop, 2022, pp. 1– 13

  11. [19]

    DiM: Distilling dataset into generative model,

    Kai Wang, Jianyang Gu, Daquan Zhou, Zheng Zhu, Wei Jiang, and Yang You, “DiM: Distilling dataset into generative model,” arXiv preprint arXiv:2303.04707, 2023

  12. [20]

    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,” inProc. CVPR Workshop, 2024, pp. 7664–7671

  13. [21]

    An efficient dataset condensation plugin and its application to continual learning,

    Enneng Yang, Li Shen, Zhenyi Wang, Tongliang Liu, and Guibing Guo, “An efficient dataset condensation plugin and its application to continual learning,” inProc. NeurIPS, 2023, pp. 67625–67642

  14. [22]

    Un- locking the potential of federated learning: The symphony of dataset distillation via deep generative latents,

    Yuqi Jia, Saeed Vahidian, Jingwei Sun, Jianyi Zhang, Vyach- eslav Kungurtsev, Neil Zhenqiang Gong, and Yiran Chen, “Un- locking the potential of federated learning: The symphony of dataset distillation via deep generative latents,” inProc. ECCV, 2024, pp. 18–33

  15. [23]

    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, 2020, pp. 305–309

  16. [24]

    Compressed gastric image generation based on soft-label dataset distillation for medical data sharing,

    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, vol. 227, pp. 107189, 2022

  17. [25]

    Image super- resolution via iterative refinement,

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi, “Image super- resolution via iterative refinement,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4713–4726, 2023

  18. [26]

    High-resolution image synthe- sis with latent diffusion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer, “High-resolution image synthe- sis with latent diffusion models,” inProc. CVPR, 2022, pp. 10684–10695

  19. [27]

    Efficient dataset distillation via minimax diffusion,

    Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen, “Efficient dataset distillation via minimax diffusion,” inProc. CVPR, 2024, pp. 15793–15803

  20. [28]

    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 Workshop, 2024, pp. 1–12

  21. [29]

    Latent dataset distillation with diffusion models,

    Brian B. Moser, Federico Raue, Sebastian Palacio, Stanislav Frolov, and Andreas Dengel, “Latent dataset distillation with diffusion models,”arXiv preprint arXiv:2403.03881, pp. 1–26, 2024

  22. [30]

    Auto-encoding varia- tional bayes,

    Kingma Diederik P. and Welling Max, “Auto-encoding varia- tional bayes,”arXiv preprint arXiv:1312.6114, pp. 1–14, 2013

  23. [31]

    Active learning for convolu- tional neural networks: A core-set approach,

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

  24. [32]

    Herding dynamical weights to learn,

    Max Welling, “Herding dynamical weights to learn,” inProc. ICML, 2009, pp. 1121–1128

  25. [33]

    Dataset condensation with distri- bution matching,

    Bo Zhao and Hakan Bilen, “Dataset condensation with distri- bution matching,” inProc. WACV, 2023, pp. 6514–6523

  26. [34]

    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, 2009, pp. 248–255

  27. [35]

    imagenette,

    Fastai, “imagenette,”https://github.com/fastai/ imagenette

  28. [36]

    Dynamic few-shot vi- sual learning without forgetting,

    Spyros Gidaris and Nikos Komodakis, “Dynamic few-shot vi- sual learning without forgetting,” inProc. CVPR, 2018, pp. 4367–4375

  29. [37]

    Deep residual learning for image recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” inProc. CVPR, 2016, pp. 770–778

  30. [38]

    Difffit: Unlocking transferability of large diffusion models via simple parameter- efficient fine-tuning,

    Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhao- qiang Liu, Jiawei Li, and Zhenguo Li, “Difffit: Unlocking transferability of large diffusion models via simple parameter- efficient fine-tuning,” inProc. ICCV, 2023, pp. 4207–4216

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.