Pith. sign in

REVIEW 4 major objections 5 minor 43 references

From Imitation to Innovation: The Emergence of AI Unique Artistic Styles and the Challenge of Copyright Protection

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that AI-generated art gains copyright exactly when its style is consistent, distinct from human styles, and prompt-faithful, and that this test can be automated with clustering plus a multimodal report generator.

desk verdict The AICD dataset and DGC clustering method are genuine contributions, but the headline 0.86/0.88 accuracy is overfit: the decision thresholds are tuned on the same AICD evaluation set that produces the final F1, with no held-out split. read the letter →

arxiv 2507.04769 v1 pith:HDCPMK4E submitted 2025-07-07 cs.CV cs.AI

Zexi Jia , Chuanwei Huang , Yeshuang Zhu , Hongyan Fei , Ying Deng , Zhiqiang Yuan , Jiapei Zhang , Jinchao Zhang
show 1 more author
Jie Zhou
This is my paper
classification cs.CVcs.AI
keywords AI-generatedartcopyrightprotectionartisticstyleclusteringmultimodallearningvision-languagemodelinterpretabilitysimilaritybenchmarkdataset
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

This paper tries to settle a question courts have not answered cleanly: when does a set of AI-generated images have its own artistic style, and therefore its own copyright? It argues that three conditions are both necessary and sufficient—consistency across generations, uniqueness relative to human artists, and accuracy relative to the prompts that produced the work—and it turns those conditions into measurable thresholds on embedding distances. To make the test usable, it builds a clustering method that uses machine-generated style descriptions alongside visual features, and an interpretable report generator that explains the verdict in plain language. On a new dataset annotated by artists and legal experts, the full system reaches higher accuracy and F1 than general-purpose multimodal models and prior style classifiers, and legal professionals rate its explanations substantially higher than those of unguided models. If the paper is right, copyright evaluation for AI art can move from case-by-case litigation to a transparent, repeatable measurement.

What carries the argument

The central machinery is Description-Guided Clustering (DGC), a style-description-based multimodal clustering method. Each artwork is embedded by a visual encoder and its automatically generated style description by a text encoder; two clustering heads assign soft labels, and a cross-modal distillation loss, a confidence loss, and a cluster-entropy loss pull the two assignments together while preventing collapse. ArtBulb then feeds the resulting cluster statistics—and the nearest human-artist clusters—to a multimodal large language model that produces a human-readable copyright report. The thresholds epsilon_c = 0.60, epsilon_d = 0.25, and epsilon_a = 0.50 are the legal judgment layer: they convert the three legal criteria into a binary verdict.

What would settle it

Run the full pipeline on a set of already-decided AI-art copyright cases and compare its binary verdict with each court's actual outcome; any case where the pipeline calls a protected work unprotected, or an infringing work non-infringing, would refute the claim that the thresholds operationalize the law. The paper's own 19 real-case images are a natural first test set.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a body of AI-generated artwork can be certified as copyrightable when it forms a tight, isolated, prompt-faithful cluster in a style-embedding space. The paper restates copyright's vague originality requirement as three inequalities: pairwise intra-cluster distance below 0.60, minimum distance to human-artist clusters above 0.25, and AMI between image-based and text-based clusterings above 0.50. It implements this as Description-Guided Clustering plus a multimodal report generator, and reports 0.86 accuracy and 0.88 F1 on the AICD benchmark, compared with 0.42 for an unguided general-purpose vision-language model and 0.77–0.80 for prior style classifiers. This is presented as evidence that interpretable, quantitative copyright judgment is achievable now.

Load-bearing premise

The load-bearing premise is that the legal concepts of originality and substantial similarity can be faithfully captured by three fixed thresholds on embedding distances and cluster alignment; if courts would not decide copyrightability that way, the whole framework is measuring style similarity rather than legal copyright.

Editorial extensions

If this is right

  • Courts and creators get a concrete, quantitative test for whether a body of AI-generated artwork constitutes its own style, making copyright disputes more predictable.
  • Because new validated AI styles are folded back into the reference database, the system improves as more AI art is certified.
  • General-purpose vision-language models alone are not sufficient for copyright judgment, since the best one reaches only 42% accuracy without clustering guidance.
  • The AICD dataset gives the research community a shared, expert-annotated benchmark for AI-art copyright tasks.
  • The description-guided clustering loss is not art-specific and can be transferred to other multimodal clustering tasks.

Reading between the lines

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

  • An implication the paper leaves implicit is that the same three criteria could be reformulated for other generative media—music, text, or video—where 'style consistency' would need a temporal or linguistic definition rather than a visual embedding distance.
  • The thresholds are calibrated on one embedding space; if a different encoder replaced the one used here, the same numeric cutoffs would likely need recalibration, so the framework's legal stability depends on fixing a standard encoder.
  • The uniqueness criterion only compares against the reference corpus; a genuinely novel AI style that resembles no known artist would still pass, but a style that resembles an artist absent from the corpus could be mislabeled as unique.
  • A testable extension would be to use the three criteria as a screening tool for copyright registration offices, flagging borderline applications for human review rather than issuing final decisions.
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 manuscript proposes ArtBulb, an interpretable framework for assessing whether AI-generated artworks exhibit a copyrightable unique style. The framework is built on three criteria distilled from legal precedents—stylistic consistency, creative uniqueness, and expressive accuracy—and operationalizes them as clustering conditions in CLIP embedding space. The authors introduce Description-Guided Clustering (DGC), a multimodal clustering method that aligns visual and textual style descriptions, and AICD, a new dataset of AI-generated artworks annotated by artists and legal experts. Experiments report that DGC improves style clustering on WikiART and four other art datasets, and that the full DGC+ArtBulb pipeline achieves 0.86 accuracy and 0.88 F1 on AICD, outperforming GPT-4o, CSD, and ARTSAVANT. The paper also reports expert ratings of explanation quality for ArtBulb's outputs.

Significance. If the empirical claims hold, the manuscript would make a useful interdisciplinary contribution: it provides a concrete benchmark (AICD), an interpretable decision pipeline for style-based copyright assessment, and a clustering method that appears to improve upon existing style clustering baselines. The attempt to connect legal criteria to measurable embedding-space conditions is timely and could help bridge the gap between the legal and computer vision communities. The expert-evaluation protocol is a valuable step toward interpretability assessment. However, the central quantitative claim currently rests on an evaluation protocol that appears to tune decision thresholds on the test set, and the legal-to-technical mapping is asserted rather than demonstrated; both issues need to be resolved before the reported gains can be trusted.

major comments (4)
  1. [Section 4.4, Table 4, and Section 3.3.2] The reported headline performance appears to be a selection statistic rather than an unbiased estimate. Section 4.4 states that the optimal F1-score of 0.88 is consistently achieved when epsilon_c=0.60, epsilon_d=0.25, and epsilon_a=0.50, and Table 2 reports the same 0.88 F1 for DGC w/ ArtBulb on AICD. Section 3.3.2 first refers to the negative samples as a 'validation set' and then says they 'form an evaluation dataset,' and no train/validation/test split for AICD is described anywhere. If the thresholds are selected on the same examples used for the final metric, the 0.86/0.88 advantage over baselines is not a valid estimate of generalization. Please provide an explicit data partition and a threshold-selection protocol that does not use the test set.
  2. [Section 4.2.2 and Section 3.2] The binary decision procedure for DGC is unspecified. The method section describes DGC as a clustering algorithm that partitions A and H into clusters, but Table 2 evaluates DGC and DGC w/ ArtBulb on a binary copyright classification task. The text does not explain how the clustering outcome—in particular, whether a cluster CA satisfies Eqs. (1)–(3)—is converted into a positive or negative judgment for each AI-generated work or set of works. Without this mapping, the comparison with classification-based baselines such as CSD and ARTSAVANT is not well-defined and the reported classification metrics cannot be reproduced.
  3. [Section 1 and Section 3.1] The mapping from legal doctrine to embedding-space thresholds is asserted rather than argued. The paper states that consistency, uniqueness, and accuracy are established 'based on legal precedents [19, 30, 40],' but Eqs. (1)–(3) define these criteria in terms of CLIP-space distances, minimum separation, and AMI scores, and no legal analysis is provided to show that courts would evaluate copyrightability in this manner. The specific threshold values (epsilon_c=0.60, epsilon_d=0.25, epsilon_a=0.50) are presented as fixed in Section 3.1 and later confirmed by the ablation in Table 4, which further indicates they are empirically chosen rather than derived from doctrine. If the legal premise is not adequately supported, the framework is better described as a style-characterization tool than a copyright-judgment tool. Please either add a substantive legal justification for the operationalization or soften the claims accordingly.
  4. [Section 4.1.2 and Section 4.2.2] The comparison with baselines may be unfair because the evaluation protocols are not symmetric. Section 4.1.2 states that CSD and ARTSAVANT are fine-tuned 'using the same data as the clustering training,' but no analogous threshold-selection procedure or held-out evaluation protocol is described for these baselines in the binary classification task. DGC's thresholds are tuned on the AICD evaluation set (as shown in Table 4), whereas the classification baselines are presumably evaluated with a single fixed decision rule. Please specify the exact training, validation, and test splits used for every method, and describe any hyperparameter or threshold selection for CSD and ARTSAVANT.
minor comments (5)
  1. [Section 4.3] The case study uses 'ArtBulm' twice; this should be corrected to 'ArtBulb.'
  2. [Table 1] The table title says 'The clustering results of the styles of real artists in the AICD dataset,' but the columns refer to separate public datasets (WikiART, Video Game, Contemporary Art, Chinese Art, Children Book). Please align the title with the dataset description in Section 3.3.1.
  3. [Table 3] The table header 'α N' with a single row mixing values for α and N is confusing; please use two separate axes or clearly label which columns correspond to α and which to N.
  4. [Section 2.1] The characterization of the Kashtanova registration [32] is imprecise: the U.S. Copyright Office's registration covered the text of the work while denying copyright for the AI-generated images. Please correct this to avoid a legally inaccurate statement.
  5. [Section 3.3.2] The manuscript does not report the final number of positive and negative examples per artwork category in AICD, nor the exact composition of the evaluation set used for Table 2. Please provide per-category counts to make the benchmark reproducible.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline result 0.88 F1 is obtained by tuning thresholds on AICD itself and then reporting the same F1; this is a fit-and-report loop, not an independent prediction.

  1. fitted input called prediction [Section 3.1 (Eqs. 1-3), Section 3.3.2, Section 4.4 (Table 4), and Table 2]
    "These are added to the validation set as negative samples. By combining these negative samples with the positive examples, we form an evaluation dataset to test the accuracy of our proposed framework and existing MLLMs. ... In this paper, we set ϵc = 0.60, ϵd = 0.25, and ϵa = 0.50. ... As shown in Table 4, the optimal F1-score of 0.88 is consistently achieved when ϵc = 0.60, ϵd = 0.25, and ϵa = 0.50. ... DGC w/ ArtBulb ... 0.86 0.88"

    The decision thresholds ϵc, ϵd, and ϵa are the parameters of the legal-criteria rules (Eqs. 1-3). Section 4.4 explicitly selects them as the values that achieve 'the optimal F1-score of 0.88' on AICD, and Table 2 then reports that same 0.88 as the method's headline performance. Section 3.3.2 calls the negative samples a 'validation set' and then says they 'form an evaluation dataset,' with no held-out split described. Thus the reported F1 is a selection statistic computed on the same examples used to choose the thresholds, so the comparison against GPT-4o, CSD, and ARTSAVANT is partly forced by construction rather than an unbiased estimate of generalization.

full rationale

The paper's core empirical claim in Table 2 is partially circular: the threshold hyperparameters in Eqs. 1-3 are tuned on the AICD evaluation set itself (Section 4.4, Table 4), and the resulting 0.88 F1 is then reported as Table 2's headline performance. Section 3.3.2 confirms the same set is called both a 'validation set' and an 'evaluation dataset,' so the reported number is a selection statistic. The DGC clustering method, the WikiArt/other style-clustering benchmarks, and the collection of AICD are independent contributions and not circular. The legal-criteria operationalization is a domain assumption rather than a circularity, and self-citations appear only in related work and are not load-bearing. No other derivation step reduces by construction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central framework depends on three fitted thresholds, two fitted loss hyperparameters, and a human-defined expert-score cutoff; none of these are derived from legal doctrine or independent benchmarks. No new physical or conceptual entities are introduced. The legal criteria themselves are the most consequential 'axiom' because the entire evaluation is built on them.

free parameters (7)
  • epsilon_c (consistency threshold) = 0.60
    Chosen from Table 4 ablation as the value maximizing F1 on the AICD test set (0.88). Not derived from legal standards.
  • epsilon_d (uniqueness threshold) = 0.25
    Chosen from Table 4 ablation as the value maximizing F1 on the AICD test set (0.88). Not derived from legal standards.
  • epsilon_a (prompt-alignment threshold) = 0.50
    Chosen from Table 4 ablation as the value maximizing F1 on the AICD test set (0.88). Not derived from legal standards.
  • alpha (entropy loss weight) = 3
    Set in Section 3.2 and ablated in Table 3; the paper reports 41.45 ACC at alpha=3 and does not report a validation split.
  • hat_N (nearest-neighbor count) = 10
    Set in Section 3.2 and ablated in Table 3 on WikiART; no separate validation split reported.
  • hat_tau (softmax temperature) = 0.1
    Set in Section 3.2; no ablation reported for this value.
  • copyright retention threshold = average score above 4.0
    Positives in AICD are defined as images with expert average score above 4.0; this cutoff is a choice that defines the test set itself.
assumptions (4)
  • domain assumption Copyright protection of AI works can be operationalized by the three criteria: consistency, uniqueness, and accuracy.
    Section 1 and Section 3.1 assert these criteria following legal precedents [14, 19, 30, 40], but no legal source is quoted that defines copyrightability in terms of embedding distances or this exact triple.
  • domain assumption Style similarity is a sufficient proxy for legal substantial similarity in copyright infringement.
    Section 1 cites [11] for 'substantial similarity', but the system measures purely visual/textual style distances, leaving out expression, idea-expression dichotomy, and other infringement elements.
  • domain assumption CLIP embedding space with the chosen thresholds captures human-perceived and legally-relevant style distinctions.
    Section 3.1 Eqs (1),(2),(3) and Section 3.2 use CLIP encoders; the paper provides no evidence that CLIP distances track the legal 'substantial similarity' judgments of courts.
  • domain assumption The expert annotations (legal experts and artists) are correct ground truth for copyright status.
    Section 3.3.2 describes the annotation protocol but reports no inter-annotator agreement, so the reliability of the labels is unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Imitation to Innovation: The Emergence of AI Unique Artistic Styles and the Challenge of Copyright Protection." pith.science (2026). https://pith.science/paper/HDCPMK4E

@misc{pith2026250704769,
  author       = {Pith},
  title        = {Pith review of: From Imitation to Innovation: The Emergence of AI Unique Artistic Styles and the Challenge of Copyright Protection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDCPMK4E}},
  note         = {Machine review of arXiv:2507.04769}
}
read the original abstract

Current legal frameworks consider AI-generated works eligible for copyright protection when they meet originality requirements and involve substantial human intellectual input. However, systematic legal standards and reliable evaluation methods for AI art copyrights are lacking. Through comprehensive analysis of legal precedents, we establish three essential criteria for determining distinctive artistic style: stylistic consistency, creative uniqueness, and expressive accuracy. To address these challenges, we introduce ArtBulb, an interpretable and quantifiable framework for AI art copyright judgment that combines a novel style description-based multimodal clustering method with multimodal large language models (MLLMs). We also present AICD, the first benchmark dataset for AI art copyright annotated by artists and legal experts. Experimental results demonstrate that ArtBulb outperforms existing models in both quantitative and qualitative evaluations. Our work aims to bridge the gap between the legal and technological communities and bring greater attention to the societal issue of AI art copyrights.

Figures

Figures reproduced from arXiv: 2507.04769 by the authors.

Figure 1
Figure 1. (a) Current laws in various countries determine whether AI-generated images qualify as protected works based on several criteria: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Description-guided Clustering (DGC) ensures consistent clustering for images with their nearest style neighbors in text space. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Our main contribution is the creation of ArtBulb, an interpretable and quantifiable framework based on legal research. ArtBulb [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Case study of four different judgment outcomes. ArtBulb can accurately determine whether AI-generated works should be [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 30 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [2]

    Taxonomy to Regulation: A (Geo)Political Taxonomy for AI Risks and Regulatory Measures in the EU AI Act

    Sinan Arda. Taxonomy to regulation: A (geo) political tax- onomy for ai risks and regulatory measures in the eu ai act. arXiv preprint arXiv:2404.11476, 2024. 2

    work page Pith review arXiv 2024

  3. [3]

    Sara Bannerman, Karen Louise Smith, Joanna Redden, Opeyemi Akanbi, Sana Maqsood, Jonathan A Obar, and Tom Streeter. Submission to the standing committee on industry and technology on bill c-27, an act to enact the consumer privacy protection act, the personal information and data pro- tection tribunal act and the artificial intelligence and data act and t...

    work page 2023

  4. [4]

    Semantic-enhanced image clustering

    Shaotian Cai, Liping Qiu, Xiaojun Chen, Qin Zhang, and Longteng Chen. Semantic-enhanced image clustering. In Proceedings of the AAAI conference on artificial intelli- gence, pages 6869–6878, 2023. 7

    work page 2023

  5. [5]

    Deep adaptive image cluster- ing

    Jianlong Chang, Lingfeng Wang, Gaofeng Meng, Shiming Xiang, and Chunhong Pan. Deep adaptive image cluster- ing. In Proceedings of the IEEE international conference on computer vision, pages 5879–5887, 2017. 7

    work page 2017

  6. [6]

    An empirical study of training self-supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9640–9649, 2021. 7

    work page 2021

  7. [7]

    Ft-shield: A watermark against unauthorized fine-tuning in text-to-image diffusion models

    Yingqian Cui, Jie Ren, Yuping Lin, Han Xu, Pengfei He, Yue Xing, Lingjuan Lyu, Wenqi Fan, Hui Liu, and Jiliang Tang. Ft-shield: A watermark against unauthorized fine-tuning in text-to-image diffusion models. ACM SIGKDD Explorations Newsletter, 26(2):76–88, 2025. 3

    work page 2025

  8. [8]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 2

    2021

Show all 43 references
  1. [9]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. In Forty-first International Conference on Mach...

    2024

  2. [10]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzy´nska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 5111–5120, 2024. 3

    2024

  3. [11]

    Goldstein on copyright

    Paul Goldstein. Goldstein on copyright. (No Title), 2005. 1

    2005

  4. [12]

    Vec- tor quantized diffusion model for text-to-image synthesis

    Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vec- tor quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 10696–10706, 2022. 2

    2022

  5. [13]

    Copyright law of the people’s republic of china.promulgated by the Standing Comm

    Chung Hua J ˆen Min Kung Han and Kuo Jud Zwo Ch’¨uan Fa. Copyright law of the people’s republic of china.promulgated by the Standing Comm. Nat’l People’s Cong, 2010. 2

    2010

  6. [14]

    China’s regulations on the at- tribution of ai-generated content: an exploration based on the open-ended approach

    Xinhang He and Pingji Shan. China’s regulations on the at- tribution of ai-generated content: an exploration based on the open-ended approach. Journal Of Intellectual Property Law and Practice, page jpae109, 2024. 1

    2024

  7. [15]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  8. [16]

    Semantic to structure: Learning structural representations for infringe- ment detection

    Chuanwei Huang, Zexi Jia, Hongyan Fei, Yeshuang Zhu, Zhiqiang Yuan, Jinchao Zhang, and Jie Zhou. Semantic to structure: Learning structural representations for infringe- ment detection. arXiv preprint arXiv:2502.07323, 2025

    2025 arXiv

  9. [17]

    Event-based semantic segmentation with posterior attention

    Zexi Jia, Kaichao You, Weihua He, Yang Tian, Yongxiang Feng, Yaoyuan Wang, Xu Jia, Yihang Lou, Jingyi Zhang, Guoqi Li, et al. Event-based semantic segmentation with posterior attention. IEEE Transactions on Image Processing, 32:1829–1842, 2023

    2023

  10. [18]

    Secret lies in color: Enhancing ai-generated images detection with color distribution anal- ysis

    Zexi Jia, Chuanwei Huang, Yeshuang Zhu, Hongyan Fei, Xi- aoyue Duan, Zhiqiang Yuan, Ying Deng, Jiapei Zhang, Jin- chao Zhang, and Jie Zhou. Secret lies in color: Enhancing ai-generated images detection with color distribution anal- ysis. In Proceedings of the Computer Vision a...

    2025

  11. [19]

    Attribution problem of generative ai: a view from us copyright law

    Zeynep ¨Ulk¨u Kahveci. Attribution problem of generative ai: a view from us copyright law. Journal of Intellectual Prop- erty Law and Practice, 18(11):796–807, 2023. 1

    2023

  12. [20]

    Ablating con- cepts in text-to-image diffusion models

    Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating con- cepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 22691–22702, 2023. 3

    2023

  13. [21]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In In- ternational conference on machine learning , pages 19730– 19742. PMLR, 2023. 7

    2023

  14. [22]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6

    2017 arXiv

  15. [23]

    Deepseek-vl: towards real-world vision- language understanding

    Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision- language understanding. arXiv preprint arXiv:2403.05525,

    arXiv

  16. [24]

    Rethinking copyright infringements in the era of text-to-image generative models

    Mazda Moayeri, Samyadeep Basu, Sriram Balasubramanian, Priyatham Kattakinda, Atoosa Chegini, Robert Brauneis, and Soheil Feizi. Rethinking copyright infringements in the era of text-to-image generative models. In The Thirteenth International Conference on Learning Representati...

  17. [25]

    Text-to-concept (and back) via cross-model align- ment

    Mazda Moayeri, Keivan Rezaei, Maziar Sanjabi, and Soheil Feizi. Text-to-concept (and back) via cross-model align- ment. In International Conference on Machine Learning , pages 25037–25060. PMLR, 2023. 3

    2023

  18. [26]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7

    2023 arXiv

  19. [27]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2

    2023 arXiv

  20. [28]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  21. [29]

    Hierarchical text-conditional image gener- ation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2

    2022 arXiv

  22. [30]

    tencent dreamwriter

    3 Regulations for the Implementation of the Copyright Law of the People’s Republic of China, Arts. 2. “tencent dreamwriter” decision of the people’s court of nanshan (dis- trict of shenzhen) 24 december 2019–case no.(2019) yue 0305 min chu no. 14010, 2020. 1

    2019

  23. [31]

    Glaze: Protecting artists from style mimicry by {Text-to-Image} models

    Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23) , pages 2187–2204, 2023. 3

    2023

  24. [32]

    Auctor ex machina: Artificial intelligence and authorship in zarya of the dawn

    Julianna Simpson. Auctor ex machina: Artificial intelligence and authorship in zarya of the dawn. Journal of Business & Technology Law, 20(1):4, 2025. 2

    2025

  25. [33]

    Diffusion art or digital forgery? investigating data replication in diffusion models

    Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6048–6058, 2023. 3

    2023

  26. [34]

    Measuring style similarity in diffusion models

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shra- may Palta, Micah Goldblum, Jonas Geiping, Abhinav Shri- vastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292 , 2024. 6, 8

    2024 arXiv

  27. [35]

    Preference rank- ing optimization for human alignment

    Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference rank- ing optimization for human alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18990– 18998, 2024. 7

    2024

  28. [36]

    Improved artgan for conditional synthesis of natural image and artwork

    Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394–409, 2019. 5

    2019

  29. [37]

    Scan: Learning to classify images without labels

    Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. InEuropean con- ference on computer vision, pages 268–285. Springer, 2020. 7

    2020

  30. [38]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 8

    2024 arXiv

  31. [39]

    Diagnosis: Detecting unautho- rized data usages in text-to-image diffusion models

    Zhenting Wang, Chen Chen, Lingjuan Lyu, Dimitris N Metaxas, and Shiqing Ma. Diagnosis: Detecting unautho- rized data usages in text-to-image diffusion models. arXiv preprint arXiv:2307.03108, 2023. 3

    2023 arXiv

  32. [40]

    On the copyright analysis of artificial intelli- gence products

    Meilin Wen. On the copyright analysis of artificial intelli- gence products. Lecture Notes in Education Psychology and Public Media, 53:133–140, 2024. 1

    2024

  33. [41]

    Unsupervised deep embedding for clustering analysis

    Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning , pages 478–487. PMLR,

  34. [42]

    Toward effective protection against diffusion-based mimicry through score distillation

    Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth Interna- tional Conference on Learning Representations, 2023. 3

    2023

  35. [43]

    Zhengyue Zhao, Jinhao Duan, Kaidi Xu, Chenan Wang, Rui Zhang, Zidong Du, Qi Guo, and Xing Hu. Can protective perturbation safeguard personal data from being exploited by stable diffusion? In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, p...

    2024

Pith tools

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