PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs

Jian Cao; Jianqing Zhang; Jie Fu; Qiang Yang; Tianyuan Zou; Yang Hua; Yang Liu

REVIEW 3 major objections 5 minor 18 references

PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs

T0 review · 3 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 PCEvolve claims that swapping Gaussian-mechanism voting for class-center contrastive scores under an exponential mechanism makes differentially private synthetic image generation work with as few as ten private images per class.

desk verdict Clean, incremental fix for PE's few-shot weakness; correct DP proof and broad experiments, but missing error bars and test-set tau tuning. read the letter →

arxiv 2506.05407 v1 pith:ALEVE364 submitted 2025-06-04 cs.CR

Jianqing Zhang , Yang Liu , Jie Fu , Yang Hua , Tianyuan Zou , Jian Cao , Qiang Yang This is my paper

classification cs.CR

keywords differentialprivacysyntheticdatagenerationfew-shotlearningexponentialmechanismcontrastivegenerativeAPIsspecializeddomainsimageclassification

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 claims that a few-shot private dataset, as small as ten images per class, can be turned into a high-utility differentially private synthetic dataset by changing how the Private Evolution loop selects prototypical images. Private Evolution uses Gaussian-mechanism similarity voting, where each private image casts one noisy vote; with few images the noise swamps the votes. PCEvolve instead averages private images into per-class centers, keeps only synthetic images that are closer to their own center than to any other center, maps those distances to bounded scores with a similarity calibrator, and feeds the scores into the Exponential Mechanism. The authors report that on four specialized datasets (COVIDx, Camelyon17, KV ASIR-f, MVTecAD-l) the resulting synthetic datasets give the best downstream Top-1 accuracy among six API-assisted baselines, and Theorem 4.1 states the full loop is $\epsilon^*$-DP. If correct, this would make privacy-preserving synthetic data practical in clinics and factories that have very few private images but can query generative APIs.

What carries the argument

The load-bearing object is the composite utility $u = h \circ g$ evaluated on class centers $\bar{\mathcal{D}}_p$ in the feature space of a pre-trained encoder $E_f$. The contrastive filter $g$ returns $1$ only for a synthetic image closer, in $\ell^2$ distance, to its own averaged class center than to every other class center; the similarity calibrator $h$ then maps that distance to $[0,1]$ via $e^{-\tau \ell}$ after min-max normalization, with $\tau$ the single hyperparameter. Since $u$ has sensitivity $1$ regardless of $K$, each Exponential Mechanism call spends the same privacy budget per selection, and sequential composition gives $\epsilon^*$-DP for the whole evolution loop.

What would settle it

Run PCEvolve with a randomly initialized or otherwise class-blind encoder as $E_f$ on a two-class few-shot dataset; if the contrastive filter then accepts nearly no synthetic images, selection becomes near-random and the reported accuracy gap over Private Evolution should disappear.

Watch

Extended reading notes

Core claim

PCEvolve's central claim is that the bottleneck in few-shot private synthetic data generation is not the generative API but the DP selection step: with $K$-shot private data, Gaussian-mechanism voting adds noise larger than the $K$ votes, making selection nearly random. The paper replaces this with a private contrastive selection rule. It aggregates the private samples per class into class centers, applies a contrastive filter $g$ that accepts a synthetic image only if it is closer to its own class center than to every other center, applies a similarity calibrator $h$ that rescales accepted distances to $[0,1]$ with sensitivity $\Delta u = 1$, and samples one prototype per class with the Exponential Mechanism. Because each Exponential Mechanism call has sensitivity $1$, the $T \times C$ calls compose to $\epsilon^*$-DP (Theorem 4.1). The paper's empirical claim is that this yields the best accuracy among six API-assisted baselines on four specialized few-shot datasets, including gains over Private Evolution of up to 5.44 accuracy points on Camelyon17, and that the gain persists across different generative APIs and downstream model architectures.

Load-bearing premise

The method assumes that distances from a pre-trained encoder make each class's averaged private images a trustworthy anchor, so that a synthetic image closer to its own class center than to all others is genuinely class-discriminative.

Editorial extensions

If this is right

Downstream models trained on PCEvolve's synthetic data can outperform models trained directly on the few-shot private data itself, by up to 8.20 accuracy points on COVIDx in the paper's experiments.
The same $\epsilon^*$ privacy budget buys more utility than Private Evolution in few-shot settings, because the added noise does not grow with the number of private images.
The generated data satisfy pure $\epsilon^*$-DP rather than approximate $(\epsilon,\delta)$-DP, and can be reused for any number of downstream tasks via the post-processing property.
With only $N=50$ synthetic images per class, PCEvolve's accuracy drops 0.95 points while Private Evolution drops 4.78 points, suggesting the method tolerates smaller API budgets.
The method carries over to different generative APIs (SD+IPA and online OpenJourney) and to different downstream backbones (ResNet, Inception, ViT).

Reading between the lines

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

The contrastive filter is effectively nearest-centroid classification in the encoder's feature space, so the method inherits whatever class separability that encoder provides; choosing $E_f$ is likely as influential as the privacy mechanism itself, even though the paper treats it as a default setting.
Because the sensitivity bound is $1$ independent of $K$, larger few-shot sets do not increase the privacy cost; a natural extension is to spend the saved budget on better center estimates or multiple prototypes per class rather than a single one.
The same contrastive-center-plus-Exponential-Mechanism loop should transfer to text or tabular domains wherever a strong encoder and a generative API exist, since nothing in the design is image-specific.
A direct adversarial test of the claimed guarantee would be a membership-inference audit on the emitted prototypes, since the honest-but-curious API provider is the stated threat model.

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.

Desk Editor's Note

read the letter

Zhi, quick read of PCEvolve. The core idea is clean: replace PE's Gaussian-mechanism voting with an exponential mechanism over contrastive scores built from averaged class centers plus a similarity calibrator. That's genuinely new relative to Lin et al. and the text-domain follow-ups, and it's the right fix for the few-shot regime where votes drown in Gaussian noise. The privacy analysis is a textbook EM composition; Theorem 4.1 is correct. I checked the stress-test worry about data-dependent normalization: it doesn't hold. The utility u is always in [0,1], so sensitivity is at most 1 no matter how the normalization denominators shift. The proof's Delta u = 1 is fine.

What the paper does well beyond the idea: the experiments are broad. Four specialized datasets, varying K and N, three API families, six downstream models, ablations, and a privacy-cost sweep. The gains over PE are consistent and often large. Ablations confirm both g and h contribute. Code is released.

Soft spots. No standard deviations or significance tests anywhere; three runs and the mean only. That's thin for claims like 64.04 vs 59.63. The tau hyperparameter is chosen per dataset on the test set (Table 4 is effectively a sweep over the evaluation set). Disclosed, but it inflates the headline numbers a bit. The PE baseline is modified (H=0) to adapt to few-shot; reasonable, but not exactly the original PE, so the comparison isn't apples-to-apples. Also, one proto per class per iteration seems restrictive, but the experiments show it works.

Overall: a useful, incremental step in API-assisted DP generation. The few-shot specialization is practically important, and the paper makes a strong case. I'd want error bars and a validation-set tau choice before trusting the exact margins, but the central argument holds. This deserves a serious referee. I'd send it to review and ask for those fixes.

Referee Report

3 major / 5 minor

Summary. The paper proposes PCEvolve, an API-assisted differentially private synthetic image generation algorithm for few-shot private datasets. It replaces the Gaussian-mechanism similarity voting of Private Evolution with an exponential mechanism whose utility combines a contrastive filter based on distances to private class centers and a similarity calibrator that normalizes distances. The paper proves an ε*-DP guarantee by composing T×C exponential mechanisms, and evaluates on four specialized datasets against six baselines, reporting consistent Top-1 accuracy improvements. The code is publicly released.

Significance. If the empirical claims hold, the contribution is useful: a training-free, API-based DP generation method that works in the few-shot regime, with a clean privacy theorem and extensive ablations. I checked the potential data-dependent-normalization concern and it does not invalidate the privacy proof: because the utility u is globally bounded in [0,1], |u(D,r)-u(D',r)| ≤ 1 for any adjacent D,D' and any r, so the exponential mechanism has sensitivity at most 1 and Theorem 4.1 is a correct textbook application of the exponential mechanism with sequential composition. The experimental design includes multiple baselines, privacy-cost curves, API variations, and a useful component ablation. The main weaknesses are in the experimental protocol: the principal baseline PE is modified, the only hyperparameter τ is selected using test-set accuracy, and no variance or significance information is reported.

major comments (3)

[Section 5.2, Table 1; Appendix A.3] The PE baseline is evaluated with the threshold H set to 0 rather than the original PE algorithm, and the justification that 'PE performs similarly for H≥0 when ε*>2' is asserted without reporting an experiment on these four few-shot datasets. Because PE is the principal baseline and the headline gains in Table 1 are measured against it, the comparison as presented does not establish superiority over the original PE. Please either run the original PE procedure on all four datasets or provide a dataset-specific ablation showing that H=0 is the correct few-shot instantiation and that the choice does not favor PCEvolve.
[Section 5.3, Table 4; Section 5.1] The only hyperparameter τ is selected based on test-set accuracy: Table 4 sweeps τ on the same test sets used for the headline numbers, and for KV ASIR-f the best τ is 20 (57.08%) rather than the default τ=10 (50.95%) reported in Table 1. Table 4 also shows large swings (e.g., MVTecAD-l from 40.70% at τ=1 to 59.26% at τ=10), so the reported improvements are sensitive to a test-set-tuned choice. Please either fix τ without reference to test labels (e.g., by a validation split or a pre-registered default) and report the resulting accuracies, or report the full Table 4 as the main result and discuss the sensitivity.
[Section 5.2, Table 1; Section 5.1] All experimental results are reported as means over three runs without standard deviations, error bars, or significance tests. This is particularly problematic for MVTecAD-l, whose test set has only 27 images: the reported gain of PCEvolve over PE is 1.85 percentage points, i.e., roughly half an image, and may be within run-to-run noise. Please report per-run results or confidence intervals for all tables and figures, and state whether the reported gains are consistent across seeds.

minor comments (5)

[Eq. (3)] When all candidates that pass the contrastive filter g have identical ℓ2 distances, ℓ_c_max = ℓ_c_min and the normalized expression in Eq. (3) is undefined; when no candidate passes g, all u values are 0 and the exponential mechanism samples uniformly. Specify a fallback (e.g., treat the denominator as 1 or draw uniformly) to make Algorithm 1 fully specified.
[Theorem 4.1] The statement 'PCEvolve satisfied ε*-DP' contains a typo; it should read 'PCEvolve satisfies ε*-DP'.
[Table 1] The abbreviations 'Came17' and 'MV AD' are used without definitions; please define them in the caption or at first occurrence.
[Eq. (1)] The strict inequality in Eq. (1) means that exact ties are assigned g=0; this tie-breaking rule should be stated explicitly.
[Section 5.2] The sentence 'PCEvolve surpasses baselines by up to 5.44%' refers specifically to Camelyon17; for MVTecAD-l the margin over PE is 1.85%, so the claim should be qualified by dataset or reported as a range.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: privacy proof is standard EM composition and empirical gains are measured against external baselines with a fixed hyperparameter.

full rationale

The derivation chain is self-contained. The privacy guarantee (Theorem 4.1) is a standard sequential composition of T×C applications of the exponential mechanism; each query's utility u=h∘g is explicitly bounded to [0,1] by Eq. (3), so the global sensitivity bound Δu≤1 holds regardless of the data-dependent normalization terms, and the claimed ε*-DP follows from Definitions 3.3 and 3.4 without importing any unverified prior result. The empirical comparison is against six external baselines implemented from published methods (PE, B, LE, RF, etc.) rather than against fitted versions of PCEvolve, and the main results use a uniform τ=10 across datasets rather than per-dataset optimal values. The contrastive filter and similarity calibrator are heuristic design choices evaluated by ablation, not predictions that are equivalent to their inputs. No load-bearing self-citations or imported uniqueness theorems appear; citations to the authors' own prior work are limited to implementation details (e.g., ResNet-18 usage). The disclosed τ-sweep in Table 4 is a hyperparameter sensitivity study and does not turn any output into a fitted value.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The central empirical claim rests on standard DP composition, on the bounded-range design of u, and on the assumption that pre-trained feature-space distances plus averaged class centers provide useful selection signals. The only fitted hyperparameter is the similarity calibrating factor tau, tuned on test accuracy in Table 4. No new physical or conceptual entities are introduced.

free parameters (1)

similarity calibrating factor tau = 10 default (20 best for KV ASIR-f in Table 4)
Scales normalized L2 distances before exponentiation in Eq. (3). Table 4 sweeps tau on test accuracy; the optimal value varies by dataset, so the default is a tuned hyperparameter rather than an input fixed by the algorithm.

assumptions (6)

standard math The exponential mechanism with utility u bounded in [0,1] has sensitivity at most 1 and is epsilon-DP.
Used in Theorem 4.1 via Definition 3.3; u values are bounded in [0,1], so the sensitivity bound holds.
standard math Sequential composition of T times C applications of an epsilon-DP mechanism yields total privacy cost epsilon* = T * C * epsilon.
Privacy accounting in Theorem 4.1, using Definition 3.4.
domain assumption Pre-trained encoder features provide a metric space where L2 distance is meaningful for class discrimination in specialized domains.
Section 4.1 computes all distances in feature space after applying Ef; the contrastive filter and similarity calibrator both rely on this.
domain assumption The API provider is honest-but-curious and only observes DP-selected public synthetic images.
Threat model in Section 4; the DP guarantee is invoked against membership inference by the provider.
domain assumption Averaging few-shot private samples into class centers yields a useful class representation for selection.
Section 4.1 step (1); the authors acknowledge that at K=1 the aggregation collapses to individual points.
domain assumption Downstream Top-1 classification accuracy on held-out test sets is a valid measure of synthetic data utility.
Section 5.1 defines the evaluation metric; all empirical comparisons depend on this proxy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs." pith.science (2026). https://pith.science/paper/ALEVE364

@misc{pith2026250605407,
  author       = {Pith},
  title        = {Pith review of: PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALEVE364}},
  note         = {Machine review of arXiv:2506.05407}
}

read the original abstract

The rise of generative APIs has fueled interest in privacy-preserving synthetic data generation. While the Private Evolution (PE) algorithm generates Differential Privacy (DP) synthetic images using diffusion model APIs, it struggles with few-shot private data due to the limitations of its DP-protected similarity voting approach. In practice, the few-shot private data challenge is particularly prevalent in specialized domains like healthcare and industry. To address this challenge, we propose a novel API-assisted algorithm, Private Contrastive Evolution (PCEvolve), which iteratively mines inherent inter-class contrastive relationships in few-shot private data beyond individual data points and seamlessly integrates them into an adapted Exponential Mechanism (EM) to optimize DP's utility in an evolution loop. We conduct extensive experiments on four specialized datasets, demonstrating that PCEvolve outperforms PE and other API-assisted baselines. These results highlight the potential of leveraging API access with private data for quality evaluation, enabling the generation of high-quality DP synthetic images and paving the way for more accessible and effective privacy-preserving generative API applications. Our code is available at https://github.com/TsingZ0/PCEvolve.

Figures

Figures reproduced from arXiv: 2506.05407 by the authors.

**Figure 2.** Illustration of our PCEvolve, whose core is the DP-protected selector. Different colors denote distinct data classes (two classes are explicitly shown, with others indicated by “. . .”). “Agg” and “Exp” denote the class center aggregation and the EM Mu (u = h ◦ g), respectively. few but useful data that are near the distribution boundary (Yang et al., 2021). To reduce bias from boundary data, we aggregate few-shot … view at source ↗

**Figure 3.** Top-1 accuracy of ResNet-18 on KVASIR-f with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Top-1 accuracy of ResNet-18 on COVIDx with [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Top-1 accuracy of various downstream models on [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Generated leather surface images w.r.t. MVAD-l for industry anomaly detection. The three rows show normal images, cut defects, and droplet defects. “Initial” denotes the initial synthetic images in PE and PCEvolve. “Private” denotes the real images from MVAD-l. Synthet…

**Figure 7.** Figure 7: The loss curves of ResNet-18 (He et al., 2016), which is retrained at each iteration of synthetic data generation for algorithm performance evaluation. We use a CLIP image encoder (Radford et al., 2021) as the encoder. C. Iterative Improvement in the Evolution Loop In …

**Figure 8.** Figure 8: Top-1 accuracy of ResNet-18 on COVIDx using four encoders. “CLIP” is short for CLIP image encoder. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 13 canonical work pages

[5]

Accessed: 2025- 01-22

URL https://platform.openai.com/docs/ guides/fine-tuning#vision. Accessed: 2025- 01-22. Pan, Z., Zhou, X., and Tian, H. Arbitrary style guidance for enhanced diffusion-based text-to-image generation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision,

work page 2025
[6]

J., and Choi, J

Seo, M., Cho, S., Lee, M., Misra, D., Choi, H., Kim, S. J., and Choi, J. Just say the name: Online continual learning with category names only via data generation.arXiv preprint arXiv:2403.10853,

arXiv
[7]

Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288,

Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y ., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288,

arXiv
[9]

Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721,

Ye, H., Zhang, J., Liu, S., Han, X., and Yang, W. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721,

arXiv
[10]

Progen: Progressive zero-shot dataset generation via in-context feedback

Ye, J., Gao, J., Wu, Z., Feng, J., Yu, T., and Kong, L. Progen: Progressive zero-shot dataset generation via in-context feedback. InFindings of the Association for Computa- tional Linguistics: EMNLP 2022,

work page 2022
[11]

Experimental Details We have included the necessary experimental details in the main body, and show more details here

12 PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs A. Experimental Details We have included the necessary experimental details in the main body, and show more details here. A.1. Image Generation APIs We consider three image generation APIs: Stable Diffusion (SD) (Rombach et al., 2022),...

work page 2022
[12]

Following PE (Lin et al., 2024), we manually implement SD on a server, providing an SD API with both text-to-image (t2i) and image-to-image (i2i) features

(OJ (online)). Following PE (Lin et al., 2024), we manually implement SD on a server, providing an SD API with both text-to-image (t2i) and image-to-image (i2i) features. Specifically, we use the pre-trained open-source SD v1.5 model from HuggingFace4 and wrap it to expose only the API interface and serve as an API server, keeping the model details hidden...

work page 2024
[13]

A DOMAIN image with LABEL

(tumor tissue patches from breast cancer metastases, two classes), and (3) KV ASIR-f9 (endoscopic images for gastrointestinal abnormal findings detection subset from KV ASIR (Pogorelov et al., 2017), three classes). In industry, we use MVTecAD-l10 (leather surface anomaly detection subset from MVTecAD (Bergmann et al., 2019), three classes). Specifically,...

work page 2017

Show all 18 references

[14]

refine this description of images to introduce rich context:

API to enhanceT . Specifically, we use an additional text prompt to enhance T with the LLaMA API: “refine this description of images to introduce rich context: ”. • RF (Samuel et al., 2024), which filters out bad t2i-generated data that closely resemble private data from diffe...

2024
[15]

We compute theσfor GM based on Theorem 3.2 given a total privacy costϵ ∗

to ensure DP while avoiding modifications to the generative API. We compute theσfor GM based on Theorem 3.2 given a total privacy costϵ ∗. (III)Using t2i and i2i APIs for image generation with DP: • PE (Lin et al., 2024), like PCEvolve, generates DP synthetic image datasets us...

2024
[16]

To balance the privacy-utility trade-off, selecting an appropriateϵ ∗ is crucial for different tasks and environments (Lin et al., 2024)

We observe that both PE and PCEvolve achieve lower accuracy with a smaller ϵ∗ and perform better with larger values, consistent with the DP literature (Dwork et al., 2014). To balance the privacy-utility trade-off, selecting an appropriateϵ ∗ is crucial for different tasks and...

2014
[17]

as the encoder. C. Iterative Improvement in the Evolution Loop In the main body, following (He et al., 2023), we train a new classification head for a pre-trained downstream model (e.g., ResNet-18) on the final synthetic dataset Ds. To demonstrate the iterative improvement in ...

2023
[18]

to evaluate the quality of the synthetic image data in the main body, considering scenarios where synthetic data is widely utilized in various downstream tasks outside the private client, and the original private dataset is typically not accessible. In specific cases where pri...

2023
[2020]

Wang, Y ., Shen, S., and Lim, B. Y . Reprompt: Automatic prompt editing to refine ai-generative art towards precise expressions. InProceedings of the 2023 CHI conference on human factors in computing systems,

2023
[2021]

H., Chen, Z., and Cao, Y

Fu, J., Hong, Y ., Ling, X., Wang, L., Ran, X., Sun, Z., Wang, W. H., Chen, Z., and Cao, Y . Differentially private federated learning: A systematic review.arXiv preprint arXiv:2405.08299, 2024a. Fu, J., Ye, Q., Hu, H., Chen, Z., Wang, L., Wang, K., and Ran, X. Dpsur: Accelera...
[2022]

A Simple Framework for Contrastive Learning of Visual Representations

Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A Simple Framework for Contrastive Learning of Visual Representations. InInternational Conference on Machine Learning (ICML), 2020a. Chen, X., Tang, S., Zhu, R., Yan, S., Jin, L., Wang, Z., Su, L., Zhang, Z., Wang, X., and T...

2024
[2023]

L., Wiles, O., and Balle, B

Ghalebikesabi, S., Berrada, L., Gowal, S., Ktena, I., Stan- forth, R., Hayes, J., De, S., Smith, S. L., Wiles, O., and Balle, B. Differentially private diffusion mod- els generate useful synthetic images.arXiv preprint arXiv:2302.13861,

arXiv
[2024]

P., and Gong, Y

Hu, R., Guo, Y ., Ratazzi, E. P., and Gong, Y . Differentially private federated learning for resource-constrained inter- net of things.arXiv preprint arXiv:2003.12705,

2003 arXiv

Pith tools

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

[1] [5]

Accessed: 2025- 01-22

URL https://platform.openai.com/docs/ guides/fine-tuning#vision. Accessed: 2025- 01-22. Pan, Z., Zhou, X., and Tian, H. Arbitrary style guidance for enhanced diffusion-based text-to-image generation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision,

work page 2025

[2] [6]

J., and Choi, J

Seo, M., Cho, S., Lee, M., Misra, D., Choi, H., Kim, S. J., and Choi, J. Just say the name: Online continual learning with category names only via data generation.arXiv preprint arXiv:2403.10853,

arXiv

[3] [7]

Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288,

Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y ., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288,

arXiv

[4] [9]

Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721,

Ye, H., Zhang, J., Liu, S., Han, X., and Yang, W. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721,

arXiv

[5] [10]

Progen: Progressive zero-shot dataset generation via in-context feedback

Ye, J., Gao, J., Wu, Z., Feng, J., Yu, T., and Kong, L. Progen: Progressive zero-shot dataset generation via in-context feedback. InFindings of the Association for Computa- tional Linguistics: EMNLP 2022,

work page 2022

[6] [11]

Experimental Details We have included the necessary experimental details in the main body, and show more details here

12 PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs A. Experimental Details We have included the necessary experimental details in the main body, and show more details here. A.1. Image Generation APIs We consider three image generation APIs: Stable Diffusion (SD) (Rombach et al., 2022),...

work page 2022

[7] [12]

Following PE (Lin et al., 2024), we manually implement SD on a server, providing an SD API with both text-to-image (t2i) and image-to-image (i2i) features

(OJ (online)). Following PE (Lin et al., 2024), we manually implement SD on a server, providing an SD API with both text-to-image (t2i) and image-to-image (i2i) features. Specifically, we use the pre-trained open-source SD v1.5 model from HuggingFace4 and wrap it to expose only the API interface and serve as an API server, keeping the model details hidden...

work page 2024

[8] [13]

A DOMAIN image with LABEL

(tumor tissue patches from breast cancer metastases, two classes), and (3) KV ASIR-f9 (endoscopic images for gastrointestinal abnormal findings detection subset from KV ASIR (Pogorelov et al., 2017), three classes). In industry, we use MVTecAD-l10 (leather surface anomaly detection subset from MVTecAD (Bergmann et al., 2019), three classes). Specifically,...

work page 2017