Pith. sign in

REVIEW 3 major objections 6 minor 26 references

ID-Card Synthetic Generation: Toward a Simulated Bona fide Dataset

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Diffusion-based generation can produce synthetic ID card images that a commercial PAD system classifies as bona fide, and adding them to training data improves a PAD system's performance, potentially reducing the need for real card samples.

desk verdict A plausible new application of diffusion-based generation to synthetic bona fide ID cards, but the COTS evaluation lacks a real-bona-fide control and the training benefit is asserted rather than shown. read the letter →

arxiv 2508.13078 v1 pith:DWJJ7EU7 submitted 2025-08-18 cs.CV

Qingwen Zeng , Juan E. Tapia , Izan Garcia , Juan M. Espin , Christoph Busch This is my paper · ORCID
classification cs.CV
keywords presentationattackdetectionIDcardssyntheticdatagenerationStableDiffusionbonafideimagesbiometricsdocumentauthenticationLoRAfine-tuning
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

Presentation-attack detection for ID cards usually trains on attack samples, but real bona fide card images are scarce and privacy-restricted. The paper tries to fill that gap by generating simulated bona fide cards: one pipeline uses Flux-1-dev one-shot diffusion to produce a full card, and another mixes a standard-compliant synthetic face with composited text, signature, and machine-readable zone layers. It reports that a commercial PAD system classifies these generated cards as bona fide, with equal-error rates around 3.6–3.8%, and that adding the simulated images to the training set improves an EfficientNet-based PAD trained from scratch. The constructive conclusion is that synthetic bona fide samples can complement or, with minimal error, replace real ID card images for training. If that holds, PAD systems can be built from less sensitive data and generalize better to unseen card types.

What carries the argument

The load-bearing mechanism is the semantic separation of face and document during generation. The hybrid pipeline creates a neutral, standard-compliant face with the HyperV AE network and then composites machine-generated text, a signature, and a machine-readable zone onto a card template; the one-shot pipeline lets a rectified-flow diffusion model (Flux-1-dev), fine-tuned with LoRA, emit the whole card in a single pass. The evaluation mechanism is a two-way probe: a commercial PAD system's classification of the synthetic images as bona fide, and a from-scratch EfficientNet PAD trained on real plus simulated samples. The first probe is meant to establish distributional closeness; the second

What would settle it

Audit the commercial PAD system's training data for synthetic images; if synthetic bona fide images were present, the test is contaminated. Alternatively, train an open-source PAD on a verified synthetic-free set and check whether the generated cards are still classified as bona fide with error rates comparable to the reported EERs around 3.6–3.8%.

Watch

Extended reading notes

Core claim

The central discovery is that the simulated bona fide distribution can be approximated well enough to matter for PAD training. Generated either with a hybrid method—a HyperV AE synthesized face combined with layer-composited text, signature, and MRZ fields—or with a one-shot Flux-1-dev diffusion generator refined by LoRA, the simulated cards are classified as bona fide by a commercial off-the-shelf PAD system. The paper interprets this acceptance as evidence that the synthetic samples lie close to the real bona fide distribution. In a from-scratch EfficientNet PAD evaluation, supplementing real bona fide images with these simulated images raises detection performance, and the paper concludes

Load-bearing premise

The claim rests on the assumption that the commercial PAD system's bona fide/attack decisions are an unbiased measure of how close the generated cards are to real ID cards, and specifically that this COTS system's training set contained no synthetic images.

Editorial extensions

If this is right

  • If the result holds, PAD training for ID cards no longer depends on collecting large sets of real card images, which are privacy-sensitive and hard to share.
  • The hybrid method's success suggests that explicit face/document compositing is a viable route to high-fidelity bona fide generation, potentially transferable to other documents such as passports and residence permits.
  • Because the COTS system labels the generated cards as bona fide, the same generation method can be read as a stress test, showing how easily a PAD system can be surprised by unseen bona fide-like inputs.
  • The reported performance gain in EfficientNet comes from a limited number of synthetic samples, so scaling and diversifying the simulated set is the direct next step implied by the paper's own conclusion.
  • If these methods are reproduced with the released configuration files, the simulated bona fide dataset can be shared with researchers without exposing real identity data.

Reading between the lines

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

  • An implication the paper does not develop: if acceptance by a COTS PAD is a reliable proxy for distributional closeness, the same synthetic pipeline could be used to build calibration or boundary tests for PAD systems, finding the least-realistic card that still counts as bona fide.
  • A testable extension: measure how PAD performance scales with the number and diversity of synthetic bona fide samples, and with the number of document templates, rather than the single template used here.
  • A caution about the central evidence: the claim that the COTS system does not contain synthetic images is an assumption about a closed system; an independent audit of its training data or a controlled study with an open PAD model of known training data would put the bona fide-classification evidence on firmer ground.
  • If the generation technique is transferred to higher-security documents such as passports and visas, the same approach could serve as a generator of both bona fide-like data and realistic attack substrates, so the boundary between defense and attack data needs active management.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes two methods for generating synthetic 'bona fide' ID-card images: a hybrid method that combines synthetic face generation with compositing of text/signature/background, and a one-shot Stable Diffusion/Flux-based method. The generated samples are evaluated using a self-trained EfficientNet PAD system and a commercial off-the-shelf (COTS) PAD system. The authors report that the COTS system classifies the synthetic cards as bona fide (EERs of 3.81%, 3.61%, and lower in Fig. 11) and claim that adding these samples to training improves PAD performance and can partially replace real bona fide data. The paper also discusses ICAO compliance and promises release of generation configurations and repositories upon acceptance.

Significance. The work addresses a real and under-explored problem: most synthetic PAD research targets attack samples, whereas the scarcity of bona fide ID-card images is a practical obstacle. Using an external COTS system as an evaluation vehicle is a valuable complement to self-trained benchmarks. If the central claim were fully established, the methods could help mitigate data restrictions in ID-card PAD. The paper also contributes two concrete generation pipelines and a comparison of diffusion-based versus hybrid approaches. However, the current evidence does not yet support the strongest claims: the COTS experiment lacks a genuine-bona-fide control, and the EfficientNet training-improvement claim is reported without sufficient experimental detail or statistical support.

major comments (3)
  1. [Section VI, Fig. 11] The COTS experiment has no real-bona-fide control. EER is computed only between simulated bona fide images and print/screen attacks. A low EER demonstrates separability from re-captured attacks, not that the simulated images lie on the real-bona-fide side of the COTS decision boundary, because print/screen attacks carry moiré, luminance, and sensor artifacts that make the discrimination easier. To support the statement that 'simulated bona fide images were detected as bona fide', the authors should add a set of real ID-card bona fide images and report score distributions (or EER/APCER/BPCER) for real-versus-attack alongside synthetic-versus-attack. Without such a control, the central interpretation of Fig. 11 is not supported.
  2. [Section VI, parenthetical assertion] The claim 'COTS systems do not contain synthetic images in the trained set as bona fide' is asserted without evidence. If the COTS training data did include synthetic images, the evaluation would be circular. The authors should either obtain a written statement from the vendor about training data or perform a probe analysis comparing COTS score distributions on synthetic images versus real bona fide images. As written, this is an unverified premise on which the main conclusion rests.
  3. [Conclusion and Section VI (EfficientNet experiments)] The claim that simulated bona fide samples 'can enhance training datasets... potentially replacing real samples with minimal error' is not backed by reported experimental detail. The manuscript does not provide the EfficientNet training protocol, dataset splits, augmentation/replacement schedule, or numerical comparisons with confidence intervals. The EER differences in Fig. 11 are also reported without error bars or significance tests. Please report the full comparison (real-only vs real+synthetic vs synthetic-only) with the protocol and uncertainty measures; otherwise this load-bearing claim cannot be evaluated.
minor comments (6)
  1. [Section VI] The notation 'HyperVAE' / 'HyperV AE' / 'hyperV AE' is used inconsistently; define a single abbreviation at first use.
  2. [Section VII] 'Minimal error' is not quantified. Specify the acceptable error threshold or report the exact performance difference.
  3. [Footnote 1] The ICAO reference URL is malformed; please provide the complete DOI or full URL for ICAO Doc 9303.
  4. [Fig. 11] The figure should label axes and show the operating point/threshold used for BPCER and APCER values; the text references these values but does not define the operating point.
  5. [Abstract] 'One of the first' should be supported by a citation or a brief comparison to prior synthetic-ID work such as [7] and [9] to avoid ambiguity.
  6. [Conclusion] The statement that all techniques are reproducible 'based on JSON files and GitHub repositories (Upon acceptance)' delays verification. Consider providing an anonymized repository or configuration files during review.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central evaluation uses an external COTS system and the generation methods are not defined in terms of the evaluation outputs.

full rationale

The paper's claimed derivation chain is empirical: synthetic ID-card images are produced with Stable Diffusion-based and hybrid pipelines, and then assessed by (i) an EfficientNet PAD trained from scratch and (ii) a third-party COTS PAD. Neither evaluation reduces to the generation procedure by construction: the synthetic images are not defined by the COTS scores, and no fitted parameter is renamed as a prediction. The COTS result is an external benchmark, so the central 'detected as bona fide' claim does not collapse into the paper's own definitions. The main caveats are support gaps rather than circularities: the unsupported assertion in Section VI that 'COTS systems do not contain synthetic images in the trained set as bona fide' is a factual assumption that could undermine the interpretation if false, and the absence of a real-bona-fide control means the reported EERs only separate synthetic images from attacks, not from genuine cards. These are correctness/validity concerns, not input-output tautologies. The self-citations (refs [9], [10], [12], [22]) document prior work by the same group, but the present pipeline relies on Stable Diffusion, Flux, LoRA, and HyperV AE, and the self-citations are not used to define away the target result or to import a uniqueness theorem. Consequently, no specific circular step can be exhibited from the text, and the appropriate circularity score is 0.

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

No explicit numerical free parameters are reported in the available text. The method relies on domain assumptions about PAD-based evaluation, the proprietary COTS training set, and the fidelity of external generative models. No new physical or conceptual entities are introduced.

assumptions (4)
  • domain assumption PAD classification outcome is a valid measure of how close a synthetic ID card is to a real bona fide card.
    The paper's central evaluation interprets the COTS system's bona fide/attack decision as evidence that the generated images resemble genuine cards (Section VI, Fig. 11). If the decision boundary is not aligned with human or ICAO fidelity, the claim weakens.
  • domain assumption The COTS PAD system was not trained on synthetic ID card images.
    The paper states 'COTS systems do not contain synthetic images in the trained set as bona fide' (Section VI), but this is an unverifiable claim about a proprietary system. If false, the evaluation becomes circular.
  • domain assumption The generated images preserve enough ID card structure (text, MRZ, signature, ICAO layout) to be treated as bona fide examples rather than unrealistic artifacts.
    The methods generate faces and documents separately and composite them (Section I, IV); the paper assumes this yields cards representative of the bona fide class.
  • domain assumption Stable Diffusion, Flux-1-dev, and HyperVAE generate faces and documents with sufficient fidelity for the PAD task.
    The entire method is built on these external generators; if their outputs contain detectable artifacts, the PAD results reflect those artifacts, not genuine bona fide fidelity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ID-Card Synthetic Generation: Toward a Simulated Bona fide Dataset." pith.science (2026). https://pith.science/paper/DWJJ7EU7

@misc{pith2026250813078,
  author       = {Pith},
  title        = {Pith review of: ID-Card Synthetic Generation: Toward a Simulated Bona fide Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DWJJ7EU7}},
  note         = {Machine review of arXiv:2508.13078}
}
read the original abstract

Nowadays, the development of a Presentation Attack Detection (PAD) system for ID cards presents a challenge due to the lack of images available to train a robust PAD system and the increase in diversity of possible attack instrument species. Today, most algorithms focus on generating attack samples and do not take into account the limited number of bona fide images. This work is one of the first to propose a method for mimicking bona fide images by generating synthetic versions of them using Stable Diffusion, which may help improve the generalisation capabilities of the detector. Furthermore, the new images generated are evaluated in a system trained from scratch and in a commercial solution. The PAD system yields an interesting result, as it identifies our images as bona fide, which has a positive impact on detection performance and data restrictions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [9]

    Synthetic ID card image generation for improving presentation attack detection,

    D. Benalcazar, J. Tapia, S. Gonzalez, and C. Busch, “Synthetic ID card image generation for improving presentation attack detection,” IEEE Transactions on Information Forensics and Security , vol. 18, pp. 1814– 1824, 2023

    2023

  2. [1]

    Analyzing and improving the image quality of StyleGAN,

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of StyleGAN,” in Proc. CVPR, 2020

    2020

  3. [2]

    Alias-free generative adversarial networks,

    T. Karras, M. Aittala, S. Laine, E. H ¨ark¨onen, J. Hellsten, J. Lehtinen, and T. Aila, “Alias-free generative adversarial networks,” in Proc. NeurIPS, 2021

    2021

  4. [3]

    MIDV-500: a dataset for identity document analysis and recognition on mobile devices in video stream,

    V . Arlazarov, K. Bulatov, T. Chernov, and V . Arlazarov, “MIDV-500: a dataset for identity document analysis and recognition on mobile devices in video stream,” Computer Optics, vol. 43, pp. 818–824, 10 2019

    2019

  5. [4]

    MIDV-2020: a comprehensive benchmark dataset for identity document analysis,

    B. Konstantin, E. E, T. Daniil, S. Natalya, C. Yulia, S. Alexander, U. S, M. Zuheng, B. Jean-Christophe, L. Muzzamil, and A. Vladimir, “MIDV-2020: a comprehensive benchmark dataset for identity document analysis,” Computer Optics, vol. 46, pp. 252–270, 03 2022

    2020

  6. [5]

    Document Live- ness Challenge DLC-2021 - part 1 (or, cg),

    E. D. Polevoy Dimitry and, Sigareva Irina and, A. Vladimir, N. Dmitry, M. Zuheng, L. Muhammad, and B. Jean-Christophe, “Document Live- ness Challenge DLC-2021 - part 1 (or, cg),” May 2022

    2021

  7. [6]

    KID34K: A dataset for online identity card fraud detection,

    E.-J. Park, S.-Y . Back, J. Kim, and S. S. Woo, “KID34K: A dataset for online identity card fraud detection,” in Proceedings of the 32nd ACM International Conf. on Information and Knowledge Management , ser. CIKM ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. 5381–5385

    2023

  8. [7]

    IDNet: A novel identity document dataset via few-shot and quality-driven synthetic data generation,

    L. Xie, Y . Wang, H. Guan, S. Nag, R. Goel, N. Swamy, Y . Yang, C. Xiao, J. Prisby, R. MacIejewski, and J. Zou, “IDNet: A novel identity document dataset via few-shot and quality-driven synthetic data generation,” in Proceedings - IEEE Intl. Conf. on Big Data, BigData . Institute of Electrical and Electronics Engineers Inc., 2024, pp. 2244– 2253

    2024

Show all 26 references
  1. [8]

    Forged presentation attack detection for ID cards on remote verification systems,

    S. Gonzalez and J. E. Tapia, “Forged presentation attack detection for ID cards on remote verification systems,” Pattern Recognition, vol. 162, p. 111352, 2025. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

    2025

  2. [10]

    Open-Set: ID card presentation attack detection using neural style transfer,

    R. P. Markham, J. M. E. L ´opez, M. Nieto-Hidalgo, and J. E. Tapia, “Open-Set: ID card presentation attack detection using neural style transfer,” IEEE Access, vol. 12, pp. 68 573–68 585, 2024

    2024

  3. [11]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) , 2017

    2017

  4. [12]

    First competition on presentation attack detection on id card,

    J. E. Tapia, N. Damer, C. Busch, J. M. Espin, J. Barrachina, A. S. Rocamora, K. Ocvirk, L. Alessio, B. Batagelj, S. Patwardhan, R. Ra- machandra, R. Mudgalgundurao, K. Raja, D. Schulz, and C. Aravena, “First competition on presentation attack detection on id card,” in IEEE Int...

    2024

  5. [13]

    Domain-agnostic document authentication against practical recapturing attacks,

    C. Chen, S. Zhang, F. Lan, and J. Huang, “Domain-agnostic document authentication against practical recapturing attacks,” IEEE Transactions on Information Forensics and Security , vol. 17, pp. 2890–2905, 2022

    2022

  6. [14]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M ¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, and R. Rombach, “Scaling rectified flow transformers for high-resolution image synthesis,” in 41st Intl. Conf. on Machine Learnin...

    2024

  7. [15]

    1.58-bit flux,

    C. Yang, C. Liu, X. Deng, D. Kim, X. Mei, X. Shen, and L.-C. Chen, “1.58-bit flux,” 2024. [Online]. Available: https://arxiv.org/abs/ 2412.18653

    2024 arXiv

  8. [16]

    Eraseanything: Enabling concept erasure in rectified flow transformers,

    D. Gao, S. Lu, S. Walters, W. Zhou, J. Chu, J. Zhang, B. Zhang, M. Jia, J. Zhao, Z. Fan et al. , “Eraseanything: Enabling concept erasure in rectified flow transformers,” ICML 2025, 2024

    2025

  9. [17]

    Variational Hyper-encoding networks,

    P. Nguyen, T. Tran, S. Gupta, S. Rana, H.-C. Dam, and S. Venkatesh, “Variational Hyper-encoding networks,” in Machine Learning and Knowledge Discovery in Databases. Research Track . Springer Intl. Publishing, 2021, pp. 100–115

    2021

  10. [18]

    MIDV-2019: Challenges of the modern mobile-based document ocr,

    K. Bulatov, D. Matalov, and V . Arlazarov, “MIDV-2019: Challenges of the modern mobile-based document ocr,” 10 2019

    2019

  11. [19]

    MIDV-Holo: A dataset for id document holo- gram detection in a video stream,

    L. Koliaskina, E. Emelianova, D. Tropin, V . Popov, K. Bulatov, D. Niko- laev, and V . Arlazarov, “MIDV-Holo: A dataset for id document holo- gram detection in a video stream,” in Intl. Conf. on Document Analysis and Recognition. Springer, 2023, pp. 486–503

    2023

  12. [20]

    Synthetic dataset of ID and travel document,

    C. Boned, M. Talarmain, N. Ghanmi, G. Chiron, S. Biswas, A. M. Awal, and O. R. Terrades, “Synthetic dataset of ID and travel document,” 2024

    2024

  13. [21]

    Guilloche detection for ID authentication: A dataset and baselines,

    M. Al-Ghadi, Z. Ming, P. Gomez-Kr ¨amer, J.-C. Burie, M. Coustaty, and N. Sidere, “Guilloche detection for ID authentication: A dataset and baselines,” in IEEE 25th Intl. Workshop on Multimedia Signal Processing (MMSP), 2023, pp. 1–6

    2023

  14. [22]

    Hybrid two-stage architecture for tampering detection of chipless ID cards,

    S. Gonzalez, A. Valenzuela, and J. Tapia, “Hybrid two-stage architecture for tampering detection of chipless ID cards,” IEEE Transactions on Biometrics, Behavior, and Identity Science , vol. 3, no. 1, pp. 89–100, 2021

    2021

  15. [23]

    Frechet inception distance (FID) for evaluating gans,

    Y . Yu, W. Zhang, and Y . Deng, “Frechet inception distance (FID) for evaluating gans,” 09 2021

    2021

  16. [24]

    Off-line handwritten signature GPDS-960 corpus,

    F. Vargas, M. Ferrer, C. Travieso, and J. Alonso, “Off-line handwritten signature GPDS-960 corpus,” in 9th Intl. Conf. on Document Analysis and Recognition (ICDAR 2007) , vol. 2. IEEE, 2007, pp. 764–768

    2007

  17. [25]

    ONOT: a high-quality ICAO-compliant synthetic mugshot dataset,

    N. Di Domenico, G. Borghi, A. Franco, D. Maltoni et al. , “ONOT: a high-quality ICAO-compliant synthetic mugshot dataset,” in The 18th IEEE Intl. Conf. on Automatic Face and Gesture Recognition (FG) , 2024, pp. 1–6

    2024

  18. [26]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition , 2022, pp. 10 684–10 695

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.