Pith. sign in

REVIEW 4 major objections 5 minor 49 references

Provably Improving Generalization of Few-Shot Models with Synthetic Data

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves a bound on test error for models trained on real and synthetic data, showing synthetic samples should be close to real ones in prediction space, and turns the bound into a loss that beats prior few-shot methods.

desk verdict Real idea and honest experiments, but the proof only covers fixed predictors, so the 'provably' claim does not hold as written. read the letter →

arxiv 2505.24190 v2 pith:ECYRABQO submitted 2025-05-30 cs.LG cs.CV

Lan-Cuong Nguyen , Quan Nguyen-Tri , Bang Tran Khanh , Dung D. Le , Long Tran-Thanh , Khoat Than This is my paper
classification cs.LGcs.CV MSC 68T05
keywords few-shotlearningsyntheticdatageneralizationbounddistributiondiscrepancylocalrobustnessprototypeimageclassificationgenerativeaugmentation
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

The paper asks when synthetic images actually help a model trained on only a few real examples, a question previously answered mostly by heuristics. It proves an upper bound on the true test error of a classifier trained on the union of real and synthetic data, where the bound is a weighted sum of a real–synthetic prediction discrepancy, local robustness terms, and empirical losses. From that bound it derives a practical loss and a two-phase algorithm that combines generator fine-tuning, K-means partitioning, and prototype-style regularization. The resulting models beat existing few-shot methods on seven of ten benchmarks in the full configuration, outperforming the strongest baseline by 0.7 points on average. If the proof holds, this is a principled answer to when and how to generate and train on synthetic data in the few-shot regime.

What carries the argument

Two definitions carry the argument. The model-based discrepancy $\bar d_h(G, S) = \frac{1}{|G||S|} \sum_{u \in G, s \in S} \|h(u) - h(s)\|$ measures how differently the model's own predictions treat the synthetic and real samples, and the local robustness $R_h(s, A \mid P) = E_{z \sim P}[\|h(z) - h(s)\| : z \in A]$ measures how stable the model output is at a point within a region. Theorem 3.3 assembles these into the test-error bound, and the proof routes through a concentration inequality (Theorem A.1) that replaces per-region empirical losses with population losses using a multinomial tail bound over region counts. The loss in Eq. (7) is the right-hand side of the bound rewritten as an objective: cross-entropy on real and synthetic data, a discrepancy regularizer that pairs real and synthetic points inside each cluster, and a robustness regularizer that pairs synthetic points inside each cluster. The partition enters through K-means on the feature space, justified by a lemma showing the robustness objectives are upper-bounded by K-means clustering objectives.

What would settle it

Train Algorithm 1 on a fixed real/synthetic pool, measure the true test loss on held-out samples from $P_0$, and compute the right-hand side of the bound using the trained model's own outputs; repeat over many random draws and seeds. If the inequality fails at a rate clearly above the claimed $\delta$, the concentration step is not valid for the data-dependent model and the provable claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 3.3: for a model $h$, a real set $S$ of $n$ i.i.d. samples from $P_0$, and a synthetic set $G$ drawn from $P_g$, partitioned into local regions $Z_i$, the population loss satisfies $F(P_0, h) \le L_h \sum_{i \in T_S} \frac{g_i}{g} [\bar d_h(G_i, S_i) + R_h(G_i, Z_i \mid P_g)] + A$ with probability at least $1 - \delta$, where $A$ collects the synthetic loss, per-region real losses, local robustness on real data, and a concentration term of order $\sqrt{K}(\frac{1}{\sqrt n} + \frac{1}{\sqrt g})$ times a logarithmic factor. Each term carries a message: $\bar d_h$ penalizes synthetic samples whose predictions differ from real ones in the same region, $R_h$ demands stable predictions within each region, and the remaining terms require $h$ to fit both data sources. The paper reads the bound as saying that a good synthetic set must be close to real samples and diverse enough to force local robustness, and that a good predictor keeps real and synthetic outputs aligned while staying stable. The practical output is the loss in Eq. (7), which realizes these terms, and Algorithm 1, which picks the partition by K-means and then minimizes that loss, with a lightweight variant that skips generator fine-tuning and uses far fewer synthetic images.

Load-bearing premise

The load-bearing premise is that the concentration bound applies to the model that training actually produces, even though that model was chosen using the same data the bound is computed on; the proof only demonstrates the bound for a model fixed before the data is seen, so if the training procedure is not accounted for, the guarantee does not cover the real predictor.

Editorial extensions

If this is right

  • A good synthetic set is one whose samples look close to real ones in the model's prediction space, not necessarily in pixel space; the $\bar d_h$ term makes this precise and gives a filtering or generation criterion.
  • Training with the Eq. (7) loss, which couples a discrepancy term and a robustness term to the classification losses, should generalize better than plain fine-tuning on the pooled data; the reported gains on seven of ten datasets are the supporting evidence.
  • Corollary 3.5 implies that what matters is the discrepancy as the model perceives it: two distributions far apart in the input space can still yield strong generalization if the model's outputs treat them as similar.
  • More synthetic samples shrink the concentration term and extend local robustness to every region holding real data, so scaling the synthetic set should keep helping; the paper's sweeps over generation counts show accuracy mostly rising with $g$.
  • The lightweight variant shows that even without fine-tuning the generator, 64 synthetic images per class combined with the regularized loss matches the strongest baseline, isolating the loss design as the main driver of the gains.

Reading between the lines

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

  • Editorial inference: the 'provable' claim is not yet established for the actual trained model, because Theorem A.1 is proven for a fixed $h$ with no union bound over the hypothesis class and no stability condition on the learning map; a careful reader should treat the bound as a design principle rather than a certificate for Algorithm 1's output.
  • Editorial inference: the discrepancy-plus-robustness pair is a transferable recipe: the same two terms give testable objectives for domain adaptation (real as source, synthetic or generated as target) and for dataset distillation, where the bound would constrain where condensed prototypes are placed.
  • Editorial inference: the method's own bound predicts where it should break — with one real sample per class, regions contain no real pairs, so local robustness is undefined and the terms shrink; the reported one-shot drop relative to the strongest baseline is consistent with that prediction.
  • Editorial inference: because the regularizers act on model outputs rather than pixels, the loss should transfer across backbone architectures with only hyperparameter changes; the ResNet-50 results provide a first partial check of that claim.
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 / 5 minor

Summary. The paper proposes a theoretical framework for few-shot image classification when training on a mix of real and synthetic data. It derives an upper bound on the population loss of a classifier h in terms of a model-based real-synthetic discrepancy, local robustness terms, empirical losses, and a concentration term. It then uses this bound to design a loss function with discrepancy and robustness regularizers, an algorithm that first clusters the data and then fine-tunes a CLIP encoder with LoRA, and a lightweight variant that avoids generator fine-tuning. Experiments on ten datasets report average improvements over baselines such as DataDream.

Significance. If the main theorem were correct, the paper would give a principled, theory-guided way to generate and train with synthetic data in the few-shot regime, and the proposed regularizers (discrepancy and robustness) would be a useful contribution. The paper also honestly reports a limitation in the 1-shot setting. However, the central theoretical claim is not established: the proof treats the classifier h and the partition as fixed, while the algorithm trains h and selects the partition using the same data. The empirical evaluation additionally relies on a single seed for the main method, so neither the theoretical nor the empirical claims are currently supported at the level asserted in the title and abstract.

major comments (4)
  1. [A.1, A.3, Theorem 3.3] Theorem A.1 is stated for "a model h learned from a dataset S," but its proof treats h as fixed: the only probabilistic element is the Bretagnolle-Huber-Carol inequality applied to multinomial cell counts, and no uniform convergence, Rademacher complexity, or stability argument appears. Because Theorem 3.3 inherits this step and is applied to the trained model h = A(S ∪ G) produced by Algorithm 1, the bound in Eq. (1) is at best a pointwise bound for each fixed h. It does not control the data-dependent classifier that minimizes the right-hand side. For a two-function class where the learned h fits S perfectly and is wrong elsewhere, the missing log|H| term is exactly what separates the claimed guarantee from a false one. This is a load-bearing gap in the "provable" claim.
  2. [4.1.1, Algorithm 1 (line 4)] The partition Γ is chosen by K-means clustering on the raw images of S ∪ G, while Theorem 3.3 requires a fixed partition of the instance space Z and the proof treats the areas Zi as fixed. The data-dependent choice of Γ introduces the same double-dipping problem as the data-dependent classifier: the concentration and robustness terms are evaluated on a partition selected using the same real and synthetic samples. No uniform bound over partitions or conditioning argument is supplied, so the theory does not justify the algorithm's partitioning step.
  3. [5, Table 1 caption] The full version reports results from a single seed ("we fixed the same seed 0 for all datasets"), while all baselines are averaged over three seeds. The average gain over the strongest baseline (DataDream_{dset}) is 0.7 percentage points, and hyperparameters (λ, λ1, λ2, number of clusters, learning rate, weight decay) are tuned per dataset on the same test sets. Without variance estimates or a statistical comparison, the claim of consistent state-of-the-art outperformance is not empirically supported.
  4. [E, Table 4] The paper itself reports that in the 1-shot setting the proposed method underperforms DataDream on all three datasets. This is a scope restriction on the central claim of "superior performance compared to state-of-the-art methods" and should be stated prominently in the main text rather than in an appendix, especially since the title and abstract emphasize few-shot performance without qualification.
minor comments (5)
  1. [Eq. (7)] The robustness term contains an unbalanced parenthesis: "∥h((g1) − h(g2)∥" should read "∥h(g1) − h(g2)∥".
  2. [5.1] The dataset list includes "FGVC Aircraft" twice; the first occurrence under "general object recognition" appears to be a mistake (likely ImageNet), and the text later refers to ImageNet without listing it as a dataset.
  3. [Theorem 3.3 and Theorem A.1] The assumption states that the loss ℓ(h, z) is Lh-Lipschitz continuous "w.r.t.h", but the proofs use Lipschitz continuity in the data argument through ∥h(s) − h(u)∥; the statement should say "w.r.t. z" or clarify the intended dependence.
  4. [4.1.2] The model optimization discussion refers to minimizing "A3 and A4" as classification losses, but in the decomposition A4 is the constant term in Eq. (5) and the empirical losses are A3 and A5; the numbering in the text does not match the definitions.
  5. [A.4] The claim that replacing distances by squared distances gives an upper bound that is equivalent to a K-means problem is only up to constants and is not precisely quantified; please state the exact inequality and the role of the constant factors in Eq. (47)-(48).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the generalization bound is a direct analytic decomposition, and the main proof gap is a missing uniform-convergence step, not a circular reduction.

full rationale

The derivation chain is self-contained. Theorem 3.3 is obtained from the algebraic decomposition in Eq. (8), the Lipschitz continuity step in Eqs. (11)-(15), and Theorem A.1, whose proof bounds F(P,h)-F(S,h) via the Bretagnolle-Huber-Carol inequality for a fixed h and fixed partition. No term in the bound is fitted from held-out test accuracy, and the algorithm's loss in Eq. (7) is exactly the regularized right-hand side of Eq. (1); minimizing that loss and then measuring accuracy on disjoint test sets is a standard ERM-plus-regularizer workflow, not a self-prediction. The only overlapping-author citation, Than et al. (2025) in Remark 3.1, supports only a side remark about O(g^{-1/2}) sample-complexity optimality; it is not load-bearing for the main theorem and is an externally published result, so it does not create circularity. What the paper does lack is a uniform-convergence argument: Theorem A.1 is stated for a model h learned from a dataset S, but its proof fixes h, and no union bound over the hypothesis class, Rademacher complexity, or stability condition is given for the data-dependent choice h = A(S∪G); similarly, Algorithm 1 selects the partition by K-means on S∪G while Theorem 3.3 treats the areas Zi as fixed. This is a soundness or correctness gap, not an instance of a conclusion reducing to its premises. Appendix E also candidly records a regime where the method underperforms (1-shot: e.g., AirC 31.1 vs 25.3 for DataDream), which is inconsistent with a self-fulfilling derivation. Verdict: no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central proof requires treating the learned model as fixed for concentration, plus Lipschitz loss assumptions. The algorithm additionally relies on several per-dataset hyperparameters and on clustering in data space being a valid proxy for clustering in prediction space.

free parameters (4)
  • lambda (real-loss weight) = 4 (1 for Stanford Cars)
    Chosen manually in Eq. (7); the theorem provides no value, and results depend on it.
  • lambda1 and lambda2 regularization weights = ratio 1/10, values from {(0.1,1),(2,20),(20,200),(50,500)} per dataset
    Searched per dataset in Appendix B; they control the discrepancy and robustness terms in the loss.
  • number of clusters K = 2 times number of classes (half for ImageNet)
    Chosen after ablation in Figure 3; the partition objective in Section 4.1.1 depends on this choice.
  • learning rate and weight decay = learning rate in {2e-4,1e-4,1e-5,1e-6}, weight decay in {1e-3,5e-4,1e-4}
    Searched per dataset during fine-tuning; affects the trained model and reported accuracy.
assumptions (5)
  • domain assumption The loss function ell is Lh-Lipschitz continuous with respect to the model output h.
    Used in the proofs of Theorem 3.3 and Theorem A.1 to replace loss differences by output distances. For softmax cross-entropy this holds locally, but the constants are not tracked.
  • standard math Concentration inequalities, specifically Bretagnolle-Huber-Carol, apply to the partition counts of S and G.
    Used in Appendix A.3 to bound the difference between population and empirical losses per region.
  • ad hoc to paper The trained classifier h is treated as fixed and independent of S union G when applying the concentration step.
    This is the load-bearing gap: Theorem A.1 says 'a model h learned from S' but the proof conditions on h; no uniform bound or stability argument is supplied.
  • domain assumption In each partition cell the empirical real and synthetic distributions approximate P0 and Pg well enough to justify K-means as the partition optimizer.
    Section 4.1.1 invokes this approximation to replace the partition objective with K-means on prediction space; in the implementation clustering is done in data space, so the connection is indirect.
  • ad hoc to paper Clustering in raw image space with K-means is an adequate proxy for the prediction-space clustering used in the proof.
    The partition optimization proof in Appendix A.4 operates on model outputs h(z), but Algorithm 1 clusters on data space to save computation. No bound connects the two.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Provably Improving Generalization of Few-Shot Models with Synthetic Data." pith.science (2026). https://pith.science/paper/ECYRABQO

@misc{pith2026250524190,
  author       = {Pith},
  title        = {Pith review of: Provably Improving Generalization of Few-Shot Models with Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECYRABQO}},
  note         = {Machine review of arXiv:2505.24190}
}
read the original abstract

Few-shot image classification remains challenging due to the scarcity of labeled training examples. Augmenting them with synthetic data has emerged as a promising way to alleviate this issue, but models trained on synthetic samples often face performance degradation due to the inherent gap between real and synthetic distributions. To address this limitation, we develop a theoretical framework that quantifies the impact of such distribution discrepancies on supervised learning, specifically in the context of image classification. More importantly, our framework suggests practical ways to generate good synthetic samples and to train a predictor with high generalization ability. Building upon this framework, we propose a novel theoretical-based algorithm that integrates prototype learning to optimize both data partitioning and model training, effectively bridging the gap between real few-shot data and synthetic data. Extensive experiments results show that our approach demonstrates superior performance compared to state-of-the-art methods, outperforming them across multiple datasets.

Figures

Figures reproduced from arXiv: 2505.24190 by the authors.

Figure 1
Figure 1. Illustration of the overall algorithm pipeline outlined in Algorithm 1. First, we generate synthetic images using the labels of real images. Subsequently, both real and synthetic images are clustered. Finally, the images are fed into the classifier. The final loss is calculated based on the model’s predictions for the samples that belong to the same cluster, thereby reducing the prediction discrepancy between real a… view at source ↗
Figure 2
Figure 2. Visualization of Discrepancy and Robustness Terms Across Settings for DTD and FGVC Aircraft Datasets. The top row shows results for the DTD dataset, with discrepancy (left) and robustness (right) terms, while the bottom row shows results for the FGVC Aircraft dataset. Accuracy values are annotated at the end of each corresponding line. this bound, thereby maximizing the performance of few-shot models. Extensive expe… view at source ↗
Figure 3
Figure 3. Results with increasing number of clusters on 4 datasets Algorithm Overview: 1. Initialization:Begin with initializing the partitions and noise vectors essential for the iterations. Each iteration is one epoch of training classifier. 2. Iterative Optimization: • Partition Optimization: Utilize the MiniBatch K-means algorithm to optimize the partition (lines 11-13). Following partition updates, recalibrate dependent … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 39 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Food-101 -- mining discriminative components with random forests

    Bossard, L., Guillaumin, M., and Van Gool, L. Food-101 -- mining discriminative components with random forests. In Fleet, D., Pajdla, T., Schiele, B., and Tuytelaars, T. (eds.), Computer Vision -- ECCV 2014, pp.\ 446--461, Cham, 2014. Springer International Publishing

    work page 2014

  3. [3]

    Describing textures in the wild

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., and Vedaldi, A. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 3606--3613, 2014

    work page 2014

  4. [4]

    da Costa, V. G. T., Dall'Asen, N., Wang, Y., Sebe, N., and Ricci, E. Diversified in-domain synthesis with efficient fine-tuning for few-shot classification. ArXiv, abs/2312.03046, 2023

    arXiv 2023

  5. [5]

    The faiss library

    Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazar'e, P.-E., Lomeli, M., Hosseini, L., and J'egou, H. The faiss library. ArXiv, abs/2401.08281, 2024

    arXiv 2024

  6. [6]

    and Liu, Y

    Gan, Z. and Liu, Y. Towards a theoretical understanding of synthetic data in LLM post-training: A reverse-bottleneck perspective. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  7. [7]

    IS SYNTHETIC DATA FROM GENERATIVE MODELS READY FOR IMAGE RECOGNITION ? In The Eleventh International Conference on Learning Representations, 2023

    He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., and QI, X. IS SYNTHETIC DATA FROM GENERATIVE MODELS READY FOR IMAGE RECOGNITION ? In The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  8. [8]

    Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification

    Helber, P., Bischke, B., Dengel, A., and Borth, D. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12 0 (7): 0 2217--2226, 2019

    work page 2019

Show all 49 references
  1. [9]

    J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lo RA : Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022

    2022

  2. [10]

    E., Gozeten, H

    Ildiz, M. E., Gozeten, H. A., Taga, E. O., Mondelli, M., and Oymak, S. High-dimensional analysis of knowledge distillation: Weak-to-strong generalization and scaling laws. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  3. [11]

    Visual prompt tuning

    Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022

    2022

  4. [12]

    U., Wasim, S

    Khattak, M. U., Wasim, S. T., Naseer, M., Khan, S., Yang, M.-H., and Khan, F. S. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 15190--15200, October 2023

    2023

  5. [13]

    M., Bader, J., Alaniz, S., Schmid, C., and Akata, Z

    Kim, J. M., Bader, J., Alaniz, S., Schmid, C., and Akata, Z. Datadream: Few-shot guided dataset generation. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part LXXI, pp.\ 252–268, Berlin, Heidelberg, 2024. Spr...

    2024

  6. [14]

    3d object representations for fine-grained categorization

    Krause, J., Stark, M., Deng, J., and Fei-Fei, L. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pp.\ 554--561, 2013

    2013

  7. [15]

    Image captions are natural prompts for text-to-image models

    Lei, S., Chen, H., Zhang, S., Zhao, B., and Tao, D. Image captions are natural prompts for text-to-image models. ArXiv, abs/2307.08526, 2023

    2023 arXiv

  8. [16]

    Caltech 101, Apr 2022

    Li, F.-F., Andreeto, M., Ranzato, M., and Perona, P. Caltech 101, Apr 2022

    2022

  9. [17]

    Promptkd: Unsupervised prompt distillation for vision-language models

    Li, Z., Li, X., Fu, X., Zhang, X., Wang, W., Chen, S., and Yang, J. Promptkd: Unsupervised prompt distillation for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 26617--26626, 2024

    2024

  10. [18]

    Gendataagent: On-the-fly dataset augmentation with synthetic data

    Li, Z., Chen, L., Andrews, J., Ba, Y., Zhang, Y., and Xiang, A. Gendataagent: On-the-fly dataset augmentation with synthetic data. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  11. [19]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019

    2019

  12. [20]

    Fine-grained visual classification of aircraft, 2013

    Maji, S., Rahtu, E., Kannala, J., Blaschko, M., and Vedaldi, A. Fine-grained visual classification of aircraft, 2013

    2013

  13. [21]

    and Zisserman, A

    Nilsback, M.-E. and Zisserman, A. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pp.\ 722--729, 2008

    2008

  14. [22]

    M., Vedaldi, A., Zisserman, A., and Jawahar, C

    Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. V. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 3498--3505, 2012

    2012

  15. [23]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021

    2021 arXiv

  16. [24]

    and Honkela, A

    R \"a is \"a , O. and Honkela, A. A bias-variance decomposition for ensembles over multiple synthetic datasets. arXiv preprint arXiv:2402.03985, 2024

    2024 arXiv

  17. [25]

    High-Resolution Image Synthesis with Latent Diffusion Models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-Resolution Image Synthesis with Latent Diffusion Models . In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 10674--10685, Los Alamitos, CA, USA, June 2022. IEEE Computer Society

    2022

  18. [26]

    C., and Fei-Fei, L

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision, 115 0 (3): 0 211–252, December 2015. ISSN 0920-5691

    2015

  19. [27]

    B., Karteek, A., Larlus, D., and Kalantidis, Y

    Sariyildiz, M. B., Karteek, A., Larlus, D., and Kalantidis, Y. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 8011--8021, 2022

    2023

  20. [28]

    D 4m: Dataset distillation via disentangled diffusion model

    Su, D., Hou, J., Gao, W., Tian, Y., and Tang, B. D 4m: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5809--5818, June 2024

    2024

  21. [29]

    Gentle local robustness implies generalization

    Than, K., Phan, D., and Vu, G. Gentle local robustness implies generalization. Machine Learning, 114 0 (6): 0 142, 2025

    2025

  22. [30]

    A bag-of-prototypes representation for dataset-level applications

    Tu, W., Deng, W., Gedeon, T., and Liang, Z. A bag-of-prototypes representation for dataset-level applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  23. [31]

    Synthetic data, real errors: how (not) to publish and use synthetic data

    Van Breugel, B., Qian, Z., and Van Der Schaar, M. Synthetic data, real errors: how (not) to publish and use synthetic data. In International Conference on Machine Learning, pp.\ 34793--34808, 2023

    2023

  24. [32]

    Prototype-based dataset comparison

    van Noord , N. Prototype-based dataset comparison. In ICCV, 2023

    2023

  25. [33]

    Cafe: Learning to condense dataset by aligning features

    Wang, K., Zhao, B., Peng, X., Zhu, Z., Yang, S., Wang, S., Huang, G., Bilen, H., Wang, X., and You, Y. Cafe: Learning to condense dataset by aligning features. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 12186--12195, 2022

    2022

  26. [34]

    Wang, T., Zhu, J.-Y., Torralba, A., and Efros, A. A. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018

    2018 arXiv

  27. [35]

    A., Oliva, A., and Torralba, A

    Xiao, J., Hays, J., Ehinger, K. A., Oliva, A., and Torralba, A. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.\ 3485--3492, 2010

    2010

  28. [36]

    Robust classification with convolutional prototype learning

    Yang, H.-M., Zhang, X.-Y., Yin, F., and Liu, C.-L. Robust classification with convolutional prototype learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3474--3482, 2018. doi:10.1109/CVPR.2018.00366

    2018

  29. [37]

    Mma: Multi-modal adapter for vision-language models

    Yang, L., Zhang, R.-Y., Wang, Y., and Xie, X. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 23826--23837, June 2024

    2024

  30. [38]

    TCP: textual-based class-aware prompt tuning for visual-language model

    Yao, H., Zhang, R., and Xu, C. TCP: textual-based class-aware prompt tuning for visual-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 , pp.\ 23438--23448. IEEE , 2024

    2024

  31. [39]

    Mmrl: Multi-modal representation learning for vision-language models

    Yao, H., Zhang, R., and Xu, C. Mmrl: Multi-modal representation learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025 . IEEE , 2025

    2025

  32. [40]

    Real-fake: Effective training data synthesis through distribution matching

    Yuan, J., Zhang, J., Sun, S., Torr, P., and Zhao, B. Real-fake: Effective training data synthesis through distribution matching. In The Twelfth International Conference on Learning Representations, 2024

    2024

  33. [41]

    J., Yoo, Y., and Choe, J

    Yun, S., Han, D., Chun, S., Oh, S. J., Yoo, Y., and Choe, J. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 6022–6031. IEEE, October 2019

    2019

  34. [42]

    N., and Lopez-Paz, D

    Zhang, H., Cisse, M., Dauphin, Y. N., and Lopez-Paz, D. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018

    2018

  35. [43]

    Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners

    Zhang, R., Hu, X., Li, B., Huang, S., Deng, H., Qiao, Y., Gao, P., and Li, H. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17...

    2023

  36. [44]

    and Bilen, H

    Zhao, B. and Bilen, H. Dataset condensation with distribution matching. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp.\ 6503--6512, 2021

    2023

  37. [45]

    and Bilen, H

    Zhao, B. and Bilen, H. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp.\ 6514--6523, January 2023

    2023

  38. [46]

    Improved distribution matching for dataset condensation

    Zhao, G., Li, G., Qin, Y., and Yu, Y. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7856--7865, 2023

    2023

  39. [47]

    Toward understanding generative data augmentation

    Zheng, C., Wu, G., and Li, C. Toward understanding generative data augmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  40. [48]

    Large language models are good prompt learners for low-shot image classification

    Zheng, Z., Wei, J., Hu, X., Zhu, H., and Nevatia, R. Large language models are good prompt learners for low-shot image classification. In CVPR, 2024

    2024

  41. [49]

    C., and Liu, Z

    Zhou, K., Yang, J., Loy, C. C., and Liu, Z. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022

    2022

Pith tools

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