Pith. sign in

REVIEW 8 minor 89 references

Selecting from a fixed synthetic image pool can match real-data performance with up to 40% fewer samples by countering generators' bias toward canonical modes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 08:11 UTC pith:G7ZLI73Z

load-bearing objection Clean generator-agnostic post-filter that actually moves the needle on synthetic-image utility; the HO-HE 1-NN idea plus partition-wise scoring is new enough and the multi-setting evidence is solid.

arxiv 2607.02637 v1 pith:G7ZLI73Z submitted 2026-07-02 cs.LG cs.AI

Post-Generation Curation of Synthetic Images via Homogeneous-Heterogeneous Splitting

Disheng Liu , Tuo Liang , Chaoda Song , Yu Yin This is my paper
classification cs.LG cs.AI
keywords synthetic data curationdata selectionhomogeneous-heterogeneous splitfidelity-diversity scoregenerative modelspost-generation filteringclassificationsemantic segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Modern generators flood synthetic pools with over-repeated canonical images of each class and under-represent harder, less typical variations. This paper shows that you can improve the downstream value of those pools without touching the generator at all: simply select a smarter subset. The method first splits each real class into a Homogeneous set of local canonical representatives and a Heterogeneous set of non-redundant variation, then scores every synthetic candidate by how well it matches the real semantics of each partition while deliberately penalizing collapse back onto the canonical modes. Across classification and segmentation benchmarks the selected subsets beat standard selection baselines and reach real-data accuracy with substantially less synthetic volume; the same scoring step still helps even after the generator has already been tuned for the target task. The practical claim is that post-generation curation is a cheap, generator-agnostic complement to better generation.

Core claim

Given only a fixed pool of synthetic images and a real reference set, a partition-conditioned fidelity-diversity score that treats Homogeneous (canonical) and Heterogeneous (non-redundant) real subsets separately produces training subsets that consistently outperform existing selection methods and can match real-data performance with up to 40% fewer synthetic samples; the same criterion remains useful as a plug-in after task-tuned generators.

What carries the argument

The Homogeneous-Heterogeneous (HO-HE) split of real classes via the directed 1-nearest-neighbor graph, together with the partition-wise fidelity-diversity score that rewards semantic alignment to each subset while penalizing drift toward canonical anchors.

Load-bearing premise

That a generic pretrained feature encoder produces a nearest-neighbor graph whose Homogeneous set is a faithful minimal cover of each real class, and that generators systematically over-produce those canonical modes relative to the Heterogeneous set.

What would settle it

On a held-out class or dataset where a strong pretrained encoder yields an HO set that does not preserve nearest-neighbor reconstruction cost, or where synthetic samples are no more similar to HO than to HE, the fidelity-diversity ranking should stop outperforming random or pure-fidelity selection under the same budget.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 8 minor

Summary. The paper studies post-generation curation of a fixed synthetic image pool for downstream utility. It argues that modern generators over-produce canonical class modes and under-represent intra-class variation, and therefore partitions each real class into a Homogeneous (HO) set of local nearest-neighbor representatives and a Heterogeneous (HE) complement via the directed 1-NN graph (Prop. 1, Sec. 3.1). Synthetic candidates are then scored with a partition-conditioned fidelity–diversity criterion (Eqs. 4–6, Alg. 1) that rewards semantic alignment while penalizing collapse toward canonical anchors. The method is generator-agnostic and requires no generator retraining. Empirically, on SVHN, CIFAR-10, Tiny-ImageNet, and ImageNet-1K, and across ResNet/EfficientNet/ViT backbones, the selected subsets outperform random, CLIP-Align, RealScore, and SBSim baselines, match real-data accuracy with substantially fewer samples (up to ~40% reduction in the plug-in setting), improve OOD robustness, and remain useful as a plug-in after inference-time intervention and JoDiffusion fine-tuning for classification and segmentation.

Significance. If the empirical claims hold, the work supplies a practical, complementary lever for synthetic-data pipelines: quality control after generation rather than only during it. The HO–HE construction is elementary but well-motivated (Prop. 1 is a clean minimal nearest-neighbor cover; the diversity score has a clear geometric reading), the evaluation is broad (four in-domain benchmarks, multiple OOD suites, two architecture families, two plug-in generator interventions, ablations on α, encoder, and budget), and the method is immediately usable on any fixed pool. Strengths that should be credited include the explicit graph-theoretic guarantee, the multi-setting plug-in results (Tabs. 2–3), and the transparent reporting of means over repeated runs. The contribution is incremental rather than foundational, but it is useful and well-supported for a data-centric ML venue.

minor comments (8)
  1. Abstract and opening paragraph: “scalable training training data” is a duplicated word; fix throughout the front matter.
  2. Fig. 1 caption and surrounding text: the figure is helpful, but the notation ℱ_HO / ℱ_HE / ℱ_syn is not defined in the caption; align with the F_HO, F_HE, F_syn used in Alg. 1 and Sec. 3.2.
  3. Alg. 1, line 8 and Eq. (5): Sp_div is written as −cos(Rp − Fp, Fsyn − Fp). Clarify whether this is evaluated against every real anchor in the partition and then aggregated (max/mean), or only against a matched nearest real; the prose and the appendix code listings are slightly inconsistent on the aggregation.
  4. Sec. 3.1 / Prop. 1: the unique-nearest-neighbor assumption is stated but not discussed under ties or near-ties in high-d feature space; a one-sentence note on how ties are broken (or that they are rare after ℓ2-normalization) would help.
  5. Fig. 6 and Tabs. 10–12: CLIP-Align is omitted from the main figure “for readability” but is the weakest baseline; either include a compact panel or explicitly point readers to the appendix tables in the figure caption.
  6. Sec. 5 / Tab. 5: feature-extraction time dominates and is shared with other selection methods; stating wall-clock on a named GPU and batch size would make the “negligible overhead” claim easier to reproduce.
  7. Appendix G code listings: the diversity term is implemented as fidelity − diversity rather than the α-weighted form of Eq. (6); note that the listings are illustrative of the unweighted case or update them to match the main equation.
  8. Related work: a short pointer to classical coreset / facility-location selection and to semantic deduplication (already cited as [17]) would better situate the 1-NN cover relative to existing data-pruning literature.

Circularity Check

0 steps flagged

No significant circularity: method definitions and Prop. 1 are independent of the downstream accuracy claims they later support.

full rationale

The paper's central claim is an empirical one: a post-generation selection rule (HO-HE 1-NN partition of real features + partition-conditioned fidelity-diversity score in Eqs. 4-6 / Alg. 1) improves downstream classifiers and segmentors relative to external baselines on held-out real test sets, and remains useful as a plug-in after task-tuned generators. Prop. 1 is a self-contained graph-theoretic statement (IHO is the unique inclusion-minimal nearest-neighbor cover under the unique-NN assumption) whose short proof appears in Appendix B.3 and does not reference any downstream accuracy number. The scoring equations are defined solely from cosine similarities to the fixed real partitions and do not embed or optimize the later-reported test accuracies; alpha is an explicit hyper-parameter that is ablated rather than fitted to the target metric. Generator HO-bias (Fig. 4) and encoder sensitivity (Fig. 9) are presented as empirical observations that motivate the design, not as premises that force the accuracy gains by construction. Self-citations are limited to public generators (EDM/EDM2), standard training recipes (timm, etc.), and ordinary baselines; none supply a uniqueness theorem or ansatz that the present results merely rename. Consequently the derivation chain does not reduce any claimed prediction or first-principles result to its own inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The method rests on standard feature-space geometry plus one free trade-off weight and the modeling choice that 1-NN in-degree identifies canonical modes. No new physical entities are postulated; the HO/HE sets are derived constructs whose utility is tested empirically.

free parameters (3)
  • alpha (fidelity-diversity trade-off) = 0.5 (default)
    Hyper-parameter in Eq. 6; default 0.5 used for main tables, ablated in Fig. 8 where optimal value shifts with generator FID and budget.
  • selection budget k / top-k per partition
    User-chosen size of the curated set; experiments sweep multiple budgets but the ranking itself depends on this cut-off.
  • feature-extractor choice = MoCo v3 (main), SigLIP best in ablation
    Encoder phi that defines the 1-NN graph and cosine scores; ablated (SigLIP, DINOv3, ViT) but still a free modeling choice that can degrade if domain-mismatched.
axioms (3)
  • domain assumption Every sample has a unique within-class nearest neighbor (used for Prop. 1 uniqueness).
    Stated explicitly before Prop. 1; ties broken arbitrarily in practice.
  • domain assumption Cosine distance on l2-normalized pretrained features is a suitable metric for semantic nearest neighbors and fidelity.
    Standard in vision; invoked throughout Sec. 3.1-3.2.
  • ad hoc to paper Generators systematically over-produce IHO modes relative to IHE (Fig. 4).
    Empirical observation that motivates the diversity penalty; not proved for all generators.
invented entities (2)
  • Homogeneous (HO) / Heterogeneous (HE) partition via directed 1-NN in-degree no independent evidence
    purpose: Provide separate canonical and non-redundant real-data anchors for scoring synthetic candidates.
    Defined by Eqs. 1-2; Prop. 1 gives the minimal-cover guarantee. Independent evidence is the empirical generator bias and downstream gains, not an external physical measurement.
  • Partition-conditioned diversity score Sp_div = -cos(Rp - Fp, Fsyn - Fp) no independent evidence
    purpose: Penalize synthetic samples that collapse back toward canonical directions while remaining near a real anchor.
    Eq. 5 and Fig. 5; geometric interpretation in Appendix B.4. Utility shown only inside this paper's experiments.

pith-pipeline@v1.1.0-grok45 · 32248 in / 2698 out tokens · 32622 ms · 2026-07-12T08:11:39.031847+00:00 · methodology

0 comments
read the original abstract

Recent generative models can produce high-quality synthetic images, offering scalable training training data for data-hungry models. Existing approaches to exploiting this potential typically involve 1) training or fine-tuning generators, or 2) using lightweight post-hoc adaptation like prompt engineering or inference-time guidance, making them generator-specific and expertise-intensive. We study a complementary question: given a fixed pool of generated images, can downstream utility be improved purely by selecting an informative subset? The answer is yes. We show that effective selection must counter a structural bias of modern generators: they tend to over-produce canonical modes of each class while underrepresenting intra-class variation. Building on this insight, we split each real class into a canonical Homogeneous (HO) subset and a non-redundant Heterogeneous (HE) subset, then score synthetic images by a fidelity-diversity criterion that rewards semantic alignment while penalizing canonical redundancy. The method is generator-agnostic and requires no retraining. Across multiple benchmarks, it consistently outperforms state-of-the-art data selection baselines and matches the real-data performance with up to 40% fewer synthetic samples. The same criterion remains effective when applied on top of stronger task-tuned generators, with gains on both classification and segmentation tasks. Post-generation selection is therefore not a substitute for better generators, but a complementary mechanism for improving the utility of synthetic data.

Figures

Figures reproduced from arXiv: 2607.02637 by Chaoda Song, Disheng Liu, Tuo Liang, Yu Yin.

Figure 1
Figure 1. Figure 1: Data selection while considering both fidelity and diversity. Real data (tar￾geted distribution)is divided into the HO￾Set and the HE-Set. Synthetic instances are then scored by referring to these two parti￾tions for subsequent selection. Aligning curated data with the target distribution is a promising post-processing principle. Image–Label Alignment methods [4, 18, 19] assume that high-quality data align… view at source ↗
Figure 2
Figure 2. Figure 2: PCA of the CIFAR-10 “horse” class under different parti￾tioning rules. The 1-NN HO-HE split (top) preserves local neigh￾borhood, whereas centroid-based split (bottom) cuts a class along a single global axis. Empirical IHO − IHE Analysis and Impact. As visualized in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: HO and HE instances in CIFAR-10. Avg. Sim. represents the average similarity between the images in each row and the entire class. HO instances are more representative, expressing the core semantics. HE instances are more diverse, capturing a broader range of variations (More examples are in the Appendix Fig.10). Top 10 Most Similar Car in HOMO-Set (Similarity = 0.886) Top 10 Least Similar Car in HETERO-Set… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the computation of S p fid and S p div in partition p. Fidelity is assessed as cos F syn i , F p i  , while diversity is measured by the angle between vectors ending at F syn i and R p i . In the diagram, Syn. instance 1 exhibits greater diversity than Syn. instance 2. 2) Diversity Score. The diversity score S p div is de￾signed to encourage the selected synthetic samples to deviate from c… view at source ↗
Figure 6
Figure 6. Figure 6: Average test accuracy over 6 runs for models trained with different amounts of selected synthetic data. We compare with RandSelect, SBSim, and RealScore; CLIP-Align is omitted for readability, with full results in Appendix Tabs. 10, 11, and 12. The black dashed line denotes the corresponding real-data baseline. 6 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: OOD Evaluation. Among different settings, [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Testing accuracy versus trade-off α (0 = fidelity, 1 = diversity). Each subfigure compares the 100K, 300K, and 500K settings with trend lines, and the optimal point on each curve is marked by a star. Results are reported with ResNet-18 using the average accuracy over 8 runs. ImageNet-Sketch [66]; 2) derivative suite, including ImageNet-C [67], -Drawing, and -Cartoon [68], derived from the validation set (A… view at source ↗
Figure 9
Figure 9. Figure 9: Testing accuracy of models trained on se [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: For CIFAR-10, we select both HO and HE images within each class. HO instances (framed in blue) are more representative and better express the core semantics of given class. HE instances (framed in green) are more diverse, capturing a broader range of variations B.3 Nearest-neighbor representative coverage For each real feature ri , define its nearest neighbor by j(i) = arg min j̸=i d(ri , rj ). The HO set… view at source ↗
Figure 11
Figure 11. Figure 11: PCA visualization in ImageNet-1K (n03769881), using different partition strategies [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The cosine similarity measures how alike instances are within the individual class in [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Performance gap between IHO and IHE across various datasets and pretrained models. For SVHN, CIFAR-10, Tiny-ImageNet, and ImageNet-1K, we evaluate multiple ResNet and ViT variants. Descriptions and links for all pretrained models are provided in [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Examples of HE instances and their most similar counterparts in HO (based on MoCo V3 features) from ImageNet-1K. Red arrows denote vectors formed by each image pair. To capture the diversity within HE, our algorithm prioritizes synthetic images that, when paired with a HE instance, yield vectors that deviate from the red-arrow directions. E Experiment Details and Full Results E.1 Generator usage in the ma… view at source ↗
Figure 15
Figure 15. Figure 15: Training recipes for different backbones. Top row: from-scratch training. Bottom row: [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Subsets of CIFAR-10-Warehouse dataset for CIFAR-10 OOD testing. [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Subsets of IN-1K OOD data Tiny-ImageNet: We utilize Tiny-ImageNet-C [64] for the evaluation of OOD, which incorporates various types of corruption. We classify them into three types: color-variation set (i.e., brightness adjustment, contrast variation), noise-variation set (i.e., pixelation, Gaussian noise, motion blur), and compression-variation set (i.e., JPEG compression). The results are presented in … view at source ↗
Figure 18
Figure 18. Figure 18: Instances with high and low CLIP-Scores in EDM generated Tiny-ImageNet. [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Instances with high and low CLIP-Scores in EDM2 generated ImageNet-1K. [PITH_FULL_IMAGE:figures/full_fig_p026_19.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 1 canonical work pages

  1. [1]

    Toward understanding generative data augmentation

    Chenyu Zheng, Guoqiang Wu, and Chongxuan Li. Toward understanding generative data augmentation. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URLhttps://openreview.net/forum?id=W5Clq1bSrR

    2023

  2. [2]

    Fake it till you make it: Learning transferable representations from synthetic imagenet clones

    Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8011–8021, 2023

    2023

  3. [3]

    Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mohammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification.Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id= DlRsoxjyPm

    2023

  4. [4]

    Is synthetic data from generative models ready for image recognition? In The Eleventh International Conference on Learning Representations, 2023

    Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In The Eleventh International Conference on Learning Representations, 2023. URL https: //openreview.net/forum?id=nUmCcZ5RKF

    2023

  5. [5]

    IS SYNTHETIC DATA USEFUL FOR TRANSFER LEARNING? AN INVESTI- GATION INTO DATA GENERATION, VOLUME, AND UTILIZATION, 2024

    Yuhang Li, Xin Dong, Chen Chen, Jingtao Li, Yuxin Wen, Michael Spranger, and Lingjuan Lyu. IS SYNTHETIC DATA USEFUL FOR TRANSFER LEARNING? AN INVESTI- GATION INTO DATA GENERATION, VOLUME, AND UTILIZATION, 2024. URL https://openreview.net/forum?id=CjPt1AC6w0

    2024

  6. [6]

    Scaling laws of synthetic images for model training

    Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training ... for now. InCVPR, pages 7382–7392, 2024. URL https://doi.org/10.1109/CVPR52733.2024.00705

    doi:10.1109/cvpr52733.2024.00705 2024

  7. [7]

    Datadream: Few-shot guided dataset generation

    Jae Myung Kim, Jessica Bader, Stephan Alaniz, Cordelia Schmid, and Zeynep Akata. Datadream: Few-shot guided dataset generation. InEuropean Conference on Computer Vision, pages 252–268. Springer, 2024

    2024

  8. [8]

    Real-fake: Effective training data synthesis through distribution matching, 2024

    Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching, 2024. URL https://arxiv.org/abs/2310. 10402

    2024

  9. [9]

    Test-time alignment of diffusion models without reward over-optimization.arXiv preprint arXiv:2501.05803, 2025

    Sunwoo Kim, Minkyu Kim, and Dongmin Park. Test-time alignment of diffusion models without reward over-optimization.arXiv preprint arXiv:2501.05803, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Diffusion curriculum: Synthetic-to-real data curriculum via image-guided diffusion

    Yijun Liang, Shweta Bhardwaj, and Tianyi Zhou. Diffusion curriculum: Synthetic-to-real data curriculum via image-guided diffusion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1697–1707, 2025

    2025

  11. [11]

    Increasing the utility of synthetic images through chamfer guidance.arXiv preprint arXiv:2508.10631, 2025

    Nicola Dall’Asen, Xiaofeng Zhang, Reyhane Askari Hemmat, Melissa Hall, Jakob Verbeek, Adriana Romero-Soriano, and Michal Drozdzal. Increasing the utility of synthetic images through chamfer guidance.arXiv preprint arXiv:2508.10631, 2025

    arXiv 2025

  12. [12]

    Shielded diffusion: Generating novel and diverse images using sparse repellency.arXiv preprint arXiv:2410.06025, 2024

    Michael Kirchhof, James Thornton, Louis Béthune, Pierre Ablin, Eugene Ndiaye, and Marco Cuturi. Shielded diffusion: Generating novel and diverse images using sparse repellency.arXiv preprint arXiv:2410.06025, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion

    Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), page 769–778. IEEE, June 2023. doi: 10.1109/cvprw59228.2023.00084. URL http://dx. ...

    doi:10.1109/cvprw59228.2023.00084 2023

  14. [14]

    Data augmentation for image classification using generative ai

    Fazle Rahat, M Shifat Hossain, Md Rubel Ahmed, Sumit Kumar Jha, and Rickard Ewetz. Data augmentation for image classification using generative ai. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4173–4182. IEEE, 2025. 10

    2025

  15. [15]

    Gonzalez, and Trevor Darrell

    Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E. Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview. net/forum?id=9wrYfqdrwk

    2023

  16. [16]

    Datacomp: In search of the next generation of multimodal datasets

    Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alex...

    2023

  17. [17]

    Semd- edup: Data-efficient learning at web-scale through semantic deduplication.arXiv preprint arXiv:2303.09540, 2023

    Amro Abbas, Kushal Tirumala, Dániel Simig, Surya Ganguli, and Ari S Morcos. Semd- edup: Data-efficient learning at web-scale through semantic deduplication.arXiv preprint arXiv:2303.09540, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Effective audio classification network based on paired inverse pyramid structure and dense mlp block

    Yunhao Chen, Yunjie Zhu, Zihui Yan, Zhen Ren, Yifan Huang, Jianlu Shen, and Lifang Chen. Effective audio classification network based on paired inverse pyramid structure and dense mlp block. InInternational Conference on Intelligent Computing, pages 70–84. Springer, 2023

    2023

  19. [19]

    Datasetgan: Efficient labeled data factory with minimal human effort

    Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145–10155, 2021

    2021

  20. [20]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In H. Wal- lach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, edi- tors,Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proc...

    2019

  21. [21]

    Explore the power of synthetic data on few-shot object detection

    Shaobo Lin, Kun Wang, Xingyu Zeng, and Rui Zhao. Explore the power of synthetic data on few-shot object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 638–647, 2023

    2023

  22. [22]

    Aerogen: Enhancing remote sensing object detection with diffusion-driven data generation

    Datao Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, Dongsheng Jiang, Yin Li, and Deyu Meng. Aerogen: Enhancing remote sensing object detection with diffusion-driven data generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 3614–3624, 2025

    2025

  23. [23]

    Noise-consistent siamese-diffusion for medical image synthesis and segmentation

    Kunpeng Qiu, Zhiqiang Gao, Zhiying Zhou, Mingjie Sun, and Yongxin Guo. Noise-consistent siamese-diffusion for medical image synthesis and segmentation. InProceedings of the Com- puter Vision and Pattern Recognition Conference, pages 15672–15681, 2025

    2025

  24. [24]

    Domain gap embeddings for generative dataset augmentation

    Yinong Oliver Wang, Younjoon Chung, Chen Henry Wu, and Fernando De la Torre. Domain gap embeddings for generative dataset augmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28684–28694, 2024

    2024

  25. [25]

    Training on thin air: Improve image classifica- tion with generated data, 2023

    Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classifica- tion with generated data, 2023. URLhttps://arxiv.org/abs/2305.15316

    Pith/arXiv arXiv 2023

  26. [26]

    Stablerep: Synthetic images from text-to-image models make strong visual representation learners, 2023

    Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners, 2023. URL https://arxiv.org/abs/2306.00984

    Pith/arXiv arXiv 2023

  27. [27]

    Synthclip: Are we ready for a fully synthetic clip training?, 2024

    Hasan Abed Al Kader Hammoud, Hani Itani, Fabio Pizzati, Philip Torr, Adel Bibi, and Bernard Ghanem. Synthclip: Are we ready for a fully synthetic clip training?, 2024. URL https: //arxiv.org/abs/2402.01832. 11

    Pith/arXiv arXiv 2024

  28. [28]

    Learning vision from models rivals learning vision from data, 2023

    Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data, 2023. URL https://arxiv. org/abs/2312.17742

    Pith/arXiv arXiv 2023

  29. [29]

    Contrastive learning with synthetic positives, 2025

    Dewen Zeng, Yawen Wu, Xinrong Hu, Xiaowei Xu, and Yiyu Shi. Contrastive learning with synthetic positives, 2025. URLhttps://arxiv.org/abs/2408.16965

    Pith/arXiv arXiv 2025

  30. [30]

    Maan Qraitem, Kate Saenko, and Bryan A. Plummer. From fake to real: Pretraining on balanced synthetic images to prevent spurious correlations in image recognition, 2024. URL https://arxiv.org/abs/2308.04553

    Pith/arXiv arXiv 2024

  31. [31]

    Will large-scale generative models corrupt future datasets? In2023 IEEE/CVF International Conference on Computer Vision (ICCV), page 20498–20508

    Ryuichiro Hataya, Han Bao, and Hiromi Arai. Will large-scale generative models corrupt future datasets? In2023 IEEE/CVF International Conference on Computer Vision (ICCV), page 20498–20508. IEEE, October 2023. doi: 10.1109/iccv51070.2023.01879. URL http: //dx.doi.org/10.1109/ICCV51070.2023.01879

    doi:10.1109/iccv51070.2023.01879 2023

  32. [32]

    Do generated data always help contrastive learning?,

    Yifei Wang, Jizhe Zhang, and Yisen Wang. Do generated data always help contrastive learning?,

  33. [33]

    URLhttps://arxiv.org/abs/2403.12448

    Pith/arXiv arXiv

  34. [34]

    Utilgen: Utility-centric generative data augmentation with dual-level task adaptation

    Jiyu Guo, Shuo Yang, Yiming Huang, Yancheng Long, Xiaobo Xia, Xiu Su, Bo Zhao, Zeke Xie, and Liqiang Nie. Utilgen: Utility-centric generative data augmentation with dual-level task adaptation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems,

  35. [35]

    URLhttps://openreview.net/forum?id=tBQEPRFT60

  36. [36]

    Effective pruning of web-scale datasets based on complexity of concept clusters

    Amro Abbas, Evgenia Rusak, Kushal Tirumala, Wieland Brendel, Kamalika Chaudhuri, and Ari S Morcos. Effective pruning of web-scale datasets based on complexity of concept clusters. arXiv preprint arXiv:2401.04578, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Learning what matters: Prioritized concept learning via relative error-driven sample selection.arXiv preprint arXiv:2506.01085, 2025

    Shivam Chandhok, Qian Yang, Oscar Mañas, Kanishk Jain, Leonid Sigal, and Aishwarya Agrawal. Learning what matters: Prioritized concept learning via relative error-driven sample selection.arXiv preprint arXiv:2506.01085, 2025

    arXiv 2025

  38. [38]

    Sampling strategies for gan synthetic data

    Binod Bhattarai, Seungryul Baek, Rumeysa Bodur, and Tae-Kyun Kim. Sampling strategies for gan synthetic data. InICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), page 2303–2307. IEEE, May 2020. doi: 10.1109/icassp40776. 2020.9054677. URLhttp://dx.doi.org/10.1109/ICASSP40776.2020.9054677

    doi:10.1109/icassp40776 2020

  39. [39]

    Datasetgan: Efficient labeled data factory with minimal human effort

    Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 10140–10150. IEEE, June 2021. doi: 10.1109/cvpr46437.2021.01001. URL http://dx.do...

    doi:10.1109/cvpr46437.2021.01001 2021

  40. [40]

    Data aug- mentation for environmental sound classification using diffusion probabilistic model with top-k selection discriminator

    Yunhao Chen, Zihui Yan, Yunjie Zhu, Zhen Ren, Jianlu Shen, and Yifan Huang. Data aug- mentation for environmental sound classification using diffusion probabilistic model with top-k selection discriminator. InInternational Conference on Intelligent Computing, pages 283–295. Springer, 2023

    2023

  41. [41]

    Strata: Self-training with task augmentation for better few-shot learning

    Tu Vu, Minh-Thang Luong, Quoc Le, Grady Simon, and Mohit Iyyer. Strata: Self-training with task augmentation for better few-shot learning. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics,

    2021

  42. [42]

    URL http://dx.doi.org/10.18653/v1/ 2021.emnlp-main.462

    doi: 10.18653/v1/2021.emnlp-main.462. URL http://dx.doi.org/10.18653/v1/ 2021.emnlp-main.462

    work page doi:10.18653/v1/2021.emnlp-main.462 2021

  43. [43]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machin...

    2021

  44. [44]

    A training-free synthetic data selection method for semantic segmentation, 2025

    Hao Tang, Siyue Yu, Jian Pang, and Bingfeng Zhang. A training-free synthetic data selection method for semantic segmentation, 2025. URLhttps://arxiv.org/abs/2501.15201. 12

    Pith/arXiv arXiv 2025

  45. [45]

    Reliable fidelity and diversity metrics for generative models, 2020

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models, 2020

    2020

  46. [46]

    Deep data augmentation for weed recognition enhancement: A diffusion probabilistic model and transfer learning based approach

    Dong Chen, Xinda Qi, Yu Zheng, Yuzhen Lu, Yanbo Huang, and Zhaojian Li. Deep data augmentation for weed recognition enhancement: A diffusion probabilistic model and transfer learning based approach. In2023 ASABE Annual International Meeting, page 1. American Society of Agricultural and Biological Engineers, 2023

    2023

  47. [47]

    Diversified in-domain synthesis with efficient fine-tuning for few-shot classification.arXiv preprint arXiv:2312.03046, 2023

    Victor G Turrisi da Costa, Nicola Dall’Asen, Yiming Wang, Nicu Sebe, and Elisa Ricci. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification.arXiv preprint arXiv:2312.03046, 2023

    Pith/arXiv arXiv 2023

  48. [48]

    Effective data augmentation with diffusion models

    Brandon Trabucco, Kyle Doherty, Max A Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=ZWzUA9zeAg

    2024

  49. [49]

    Feedback-guided data synthesis for imbalanced classification, 2024

    Reyhane Askari Hemmat, Mohammad Pezeshki, Florian Bordes, Michal Drozdzal, and Adriana Romero-Soriano. Feedback-guided data synthesis for imbalanced classification, 2024. URL https://arxiv.org/abs/2310.00158

    Pith/arXiv arXiv 2024

  50. [50]

    An empirical study of training self-supervised vision transformers.arXiv preprint arXiv:2104.02057, 2021

    Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers.arXiv preprint arXiv:2104.02057, 2021

    Pith/arXiv arXiv 2021

  51. [51]

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y . Ng. Reading digits in natural images with unsupervised feature learning. InNIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. URL http://ufldl.stanford. edu/housenumbers/nips2011_housenumbers.pdf

    2011

  52. [52]

    Cifar-10 (canadian institute for advanced research).URL http://www

    Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research).URL http://www. cs. toronto. edu/kriz/cifar. html, 5(4):1, 2010

    2010

  53. [53]

    Ya Le and Xuan S. Yang. Tiny imagenet visual recognition challenge. 2015. URL https: //api.semanticscholar.org/CorpusID:16664790

    2015

  54. [54]

    Better diffusion models further improve adversarial training

    Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. InInternational Conference on Machine Learning (ICML), 2023

    2023

  55. [55]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. InProc. CVPR, 2024

    2024

  56. [56]

    Guiding a diffusion model with a bad version of itself

    Tero Karras, Miika Aittala, Tuomas Kynkäänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. InProc. NeurIPS, 2024

    2024

  57. [57]

    Contrastive multiview coding

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. InEuropean conference on computer vision, pages 776–794. Springer, 2020

    2020

  58. [58]

    Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective.International Journal of Computer Vision, 111(1):98–136, 2015

    2015

  59. [59]

    Cashman, and Jamie Shotton

    Erroll Wood, Tadas Baltrušaitis, Charlie Hewitt, Sebastian Dziadzio, Matthew Johnson, Virginia Estellers, Thomas J. Cashman, and Jamie Shotton. Fake it till you make it: Face analysis in the wild using synthetic data alone, 2021. URLhttps://arxiv.org/abs/2109.15102

    Pith/arXiv arXiv 2021

  60. [60]

    Extracting training data from diffusion models

    Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In32nd USENIX Security Symposium (USENIX Security 23), pages 5253–5270, 2023

    2023

  61. [61]

    Dcface: Synthetic face generation with dual condition diffusion model

    Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. Dcface: Synthetic face generation with dual condition diffusion model. InProceedings of the ieee/cvf conference on computer vision and pattern recognition, pages 12715–12725, 2023. 13

    2023

  62. [62]

    Improving geo-diversity of generated images with contextualized vendi score guidance

    Reyhane Askari Hemmat, Melissa Hall, Alicia Sun, Candace Ross, Michal Drozdzal, and Adriana Romero-Soriano. Improving geo-diversity of generated images with contextualized vendi score guidance. InEuropean Conference on Computer Vision, pages 213–229. Springer, 2024

    2024

  63. [63]

    Jodiffusion: Jointly diffusing image with pixel-level annotations for semantic segmentation promotion

    Haoyu Wang, Lei Zhang, Wenrui Liu, Dengyang Jiang, Wei Wei, and Chen Ding. Jodiffusion: Jointly diffusing image with pixel-level annotations for semantic segmentation promotion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 9775–9783, 2026

    2026

  64. [64]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  65. [65]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  66. [66]

    Visual transformers: Token-based image representation and processing for computer vision, 2020

    Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision, 2020

    2020

  67. [67]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. InInternational Conference on Learning Representations, 2019. URLhttps://openreview.net/forum?id=HJz6tiCqYm

    2019

  68. [68]

    Do imagenet classifiers generalize to imagenet?, 2019

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet?, 2019. URLhttps://arxiv.org/abs/1902.10811

    Pith/arXiv arXiv 2019

  69. [69]

    Learning robust global represen- tations by penalizing local predictive power

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global represen- tations by penalizing local predictive power. InAdvances in Neural Information Processing Systems, pages 10506–10518, 2019

    2019

  70. [70]

    Benchmarking neural network robustness to common corruptions and perturbations.Proceedings of the International Conference on Learning Representations, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations.Proceedings of the International Conference on Learning Representations, 2019

    2019

  71. [71]

    Imagenet-cartoon and imagenet-drawing: two domain shift datasets for imagenet

    Tiago Salvador and Adam M Oberman. Imagenet-cartoon and imagenet-drawing: two domain shift datasets for imagenet. InICML 2022 Shift Happens Workshop, 2022. URL https: //openreview.net/forum?id=YlAUXhjwaQt

    2022

  72. [72]

    Cifar- 10-warehouse: Broad and more realistic testbeds in model generalization analysis

    Xiaoxiao Sun, Xingjian Leng, Zijian Wang, Yang Yang, Zi Huang, and Liang Zheng. Cifar- 10-warehouse: Broad and more realistic testbeds in model generalization analysis. InICLR, 2024

    2024

  73. [73]

    Encoder-decoder with atrous separable convolution for semantic image segmentation

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801–818, 2018

    2018

  74. [74]

    Training generative adversarial networks with limited data

    Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. InProc. NeurIPS, 2020

    2020

  75. [75]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. InProc. NeurIPS, 2022

    2022

  76. [76]

    Sigmoid loss for language image pre-training, 2023

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023

    2023

  77. [77]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    2025

  78. [78]

    near” set contains canonical patterns, while the “far

    Ross Wightman. Pytorch image models. https://github.com/rwightman/ pytorch-image-models, 2019. 14 Appendix The supplementary material is organized by the role each section plays in supporting the main paper. Appendix A first gives visual examples and intra-class statistics that characterize the HO/HEpartition. Appendix B explains the motivation and theore...

    arXiv 2019

  79. [79]

    Such property implies that learning the pattern in HO, and then we can reconstruct the whole original feature space with the smallest cost

    To reach any node b in the graph, starting from a node in HO(different from b) consistently yields the minimal cost. Such property implies that learning the pattern in HO, and then we can reconstruct the whole original feature space with the smallest cost

  80. [80]

    $NUM_GPU

    Suppose, in the original graph, each node has a single nearest neighbor. Then, theoretically, HOis a minimal subset of nodes such that every node in the graph can be reached from some node in HOwith minimal cost. Training downstream models on a given dataset can be interpreted as optimizing the model to capture the full set of patterns in the underlying f...

Showing first 80 references.