pith. sign in

arxiv: 2511.09749 · v2 · submitted 2025-11-12 · 💻 cs.CV · cs.LG

Gradient-Guided Exploration of Generative Model's Latent Space for Controlled Iris Image Augmentations

Mahsa Mitcheff , Siamul Karim Khan , Adam Czajka This is my paper

Pith reviewed 2026-05-17 22:57 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords iris image augmentationlatent space traversalgradient guidancegenerative adversarial networksGAN inversioniris recognitionpresentation attack detection
0
0 comments X

The pith

Gradient-guided steps through a generative model's latent space let users change specific iris features like pupil size or sharpness while keeping the same identity.

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

The paper introduces a method to augment iris images by projecting a starting image into the latent space of a pre-trained generative model and then moving the latent code in the direction given by the gradient of a chosen feature such as sharpness, pupil size, iris size, or pupil-to-iris ratio. This movement produces a new image that alters only the targeted property and leaves the identity unchanged. A sympathetic reader would care because reliable iris recognition and presentation attack detection need large datasets that show realistic variations in iris appearance, yet the complex texture of irises makes controlled synthesis difficult with existing techniques. The approach works from either randomly generated images or real iris images inverted into the latent space and extends to any attribute that can be expressed with a differentiable loss.

Core claim

The paper claims that traversing a generative model's latent space toward codes that represent same-identity samples, with the direction set by the gradient of geometrical, textural, or quality-related iris image features, yields augmented images that manipulate the desired properties while preserving identity. The strategy applies to both randomly generated images from a pre-trained GAN and real-world iris images projected via GAN inversion, and it can be extended to any attribute for which a differentiable loss term can be written.

What carries the argument

Gradient-guided latent space traversal, which computes the gradient of a chosen iris feature loss and steps the latent code along that direction to alter only the targeted attribute while holding identity fixed.

If this is right

  • Augmented images with controlled changes in sharpness, pupil size, iris size, or pupil-to-iris ratio become available for training iris recognition and presentation attack detection systems.
  • The same traversal procedure can be applied to any new attribute once a differentiable loss for that attribute is defined.
  • Both synthetic images sampled from the GAN and real iris images inverted into the latent space serve as valid starting points for augmentation.
  • Datasets can be expanded without collecting additional real iris samples while maintaining identity labels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The technique could be tested by measuring whether augmented images improve the generalization of iris matchers on held-out real data.
  • Similar gradient-guided editing might transfer to other fine-textured biometrics such as fingerprints or periocular regions.
  • If the method succeeds, it would reduce reliance on large-scale real iris collection campaigns for developing new recognition algorithms.

Load-bearing premise

Moving along the gradient in the latent space of a pre-trained generative model will reliably change only the selected iris attribute and leave the underlying identity unchanged without creating artifacts.

What would settle it

An iris recognition system matching original images against their gradient-augmented versions and showing match scores falling below the decision threshold on a substantial fraction of pairs would indicate that identity is not preserved.

Figures

Figures reproduced from arXiv: 2511.09749 by Adam Czajka, Mahsa Mitcheff, Siamul Karim Khan.

Figure 1
Figure 1. Figure 1: We manipulate selected geometrical or textural iris image attribute by traversing the latent space of a generative decoder (trained [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison score distributions between initial iris im [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between synthetic iris samples generated with and without use of identity loss component: (a) the initial iris sample., [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the iris image attribute manipulation process through gradient-guided traversal of the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Same as in Figure [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

Developing reliable iris recognition and presentation attack detection methods requires diverse datasets that capture realistic variations in iris features and a wide spectrum of anomalies. Because of the rich texture of iris images, which spans a wide range of spatial frequencies, synthesizing same-identity iris images while controlling specific attributes remains challenging. In this work, we introduce a new iris image augmentation strategy by traversing a generative model's latent space toward latent codes that represent same-identity samples but with some desired iris image properties manipulated. The latent space traversal is guided by a gradient of specific geometrical, textural, or quality-related iris image features (e.g., sharpness, pupil size, iris size, or pupil-to-iris ratio) and preserves the identity represented by the image being manipulated. The proposed approach can be easily extended to manipulate any attribute for which a differentiable loss term can be formulated. Additionally, our approach can use either randomly generated images using either a pre-train GAN model or real-world iris images. We can utilize GAN inversion to project any given iris image into the latent space and obtain its corresponding latent code.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper claims to introduce a gradient-guided traversal technique in the latent space of a pre-trained generative model (such as a GAN) to augment iris images. Starting from either randomly generated or real iris images (via GAN inversion), the method uses gradients of differentiable losses on attributes including sharpness, pupil size, iris size, and pupil-to-iris ratio to manipulate targeted geometrical, textural, or quality-related features while preserving the identity of the source image. The approach is presented as extensible to any attribute admitting a differentiable loss term.

Significance. If the central claim holds, the technique would offer a practical way to synthesize controlled variations in iris datasets for recognition and presentation attack detection tasks, addressing the challenge of rich high-frequency iris textures. The extensibility to arbitrary differentiable losses and support for both synthetic and inverted real images are positive aspects that could broaden utility in data augmentation pipelines.

major comments (2)
  1. [Abstract] Abstract: The central claim that gradient-guided latent traversal reliably preserves identity while altering only the targeted attribute is load-bearing but unsupported. No quantitative results (e.g., iris recognition matching scores, identity similarity metrics, or artifact quantification) or error analysis are provided to demonstrate that same-identity samples remain stable after traversal or that unintended shifts in other features are bounded.
  2. [Abstract] Abstract: The method assumes that the latent space of the pre-trained generative model is sufficiently disentangled with respect to high-frequency iris textures so that gradient descent on an attribute loss (e.g., pupil-to-iris ratio) isolates the desired change without identity-altering artifacts or GAN-inversion side effects. No justification, bounds, or empirical checks for this disentanglement are given, leaving the reliability of controlled augmentation unverified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the method's potential utility. We acknowledge that the claims regarding reliable identity preservation and latent space disentanglement require stronger empirical support. We will revise the manuscript to address these points through added quantitative evaluations and discussions, as detailed below.

read point-by-point responses

  1. Referee: The central claim that gradient-guided latent traversal reliably preserves identity while altering only the targeted attribute is load-bearing but unsupported. No quantitative results (e.g., iris recognition matching scores, identity similarity metrics, or artifact quantification) or error analysis are provided to demonstrate that same-identity samples remain stable after traversal or that unintended shifts in other features are bounded.

    Authors: We agree that quantitative validation is essential to substantiate the identity preservation claim. The current manuscript focuses on the methodological description and qualitative demonstrations, but we recognize this leaves the central claim insufficiently supported. In the revised version, we will add a new experimental subsection that includes: iris recognition matching scores computed with established models (e.g., reporting verification rates or similarity scores before and after traversal); embedding-based identity similarity metrics; and quantification of unintended shifts by tracking changes in non-targeted attributes along with artifact analysis. Error bars and statistics across multiple starting images and traversal steps will also be reported to bound the effects. revision: yes

  2. Referee: The method assumes that the latent space of the pre-trained generative model is sufficiently disentangled with respect to high-frequency iris textures so that gradient descent on an attribute loss (e.g., pupil-to-iris ratio) isolates the desired change without identity-altering artifacts or GAN-inversion side effects. No justification, bounds, or empirical checks for this disentanglement are given, leaving the reliability of controlled augmentation unverified.

    Authors: We concur that explicit justification and checks for the disentanglement assumption are needed, particularly given the high-frequency nature of iris textures. While the method builds on observed behaviors in latent editing literature, we did not provide iris-specific empirical verification. In the revision, we will add empirical checks by measuring post-traversal changes in identity embeddings and texture statistics (e.g., frequency-domain analysis), along with a sensitivity study on GAN inversion effects. We will also include a limitations discussion noting that full theoretical bounds on disentanglement are challenging but that the gradient-guided approach empirically isolates changes for the tested attributes. revision: yes

Circularity Check

0 steps flagged

No circularity: standard gradient descent on pre-trained GAN with external differentiable losses

full rationale

The paper's core method traverses the latent space of a pre-trained generative model using gradients of externally defined iris features (sharpness, pupil-to-iris ratio, etc.). These losses are formulated independently of the traversal process itself and do not rely on fitting parameters to the output data or self-referential definitions. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked to justify the central claim. The derivation remains self-contained, drawing on standard optimization and pre-trained models without reducing predictions to inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The approach depends on the pre-trained generative model having a latent space structured enough for identity-preserving paths and on the ability to define differentiable losses for the target attributes; these are domain assumptions rather than derived results.

free parameters (1)
  • Traversal step size or number of steps
    Must be chosen to balance attribute change against identity preservation; not derived from first principles.
axioms (2)
  • domain assumption A differentiable loss term can be formulated for any desired iris attribute (sharpness, pupil size, etc.).
    Invoked to enable gradient guidance; stated as an extension point in the abstract.
  • domain assumption Moving along the gradient in latent space alters the target attribute while approximately preserving identity.
    Central premise for the augmentation to produce useful same-identity samples.

pith-pipeline@v0.9.0 · 5495 in / 1283 out tokens · 35524 ms · 2026-05-17T22:57:32.783831+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · 1 internal anchor

  1. [1]

    gov/IREX10/

    IREX 10: Identification Track.https://pages.nist. gov/IREX10/. Accessed: Sept. 17, 2025. 6

    work page 2025

  2. [2]

    Accessed: Sept

    IREX V Homepage.https://www.nist.gov/itl/ iad/image-group/irex-v-homepage. Accessed: Sept. 17, 2025. 8

    work page 2025

  3. [3]

    com / CVRL / OpenSourceIrisRecognition/

    University of Notre Dame Open Source Iris Recogni- tion Repository.https : / / github . com / CVRL / OpenSourceIrisRecognition/. Accessed: Sept. 17,

    work page

  4. [4]

    Comprehensive study in open- set iris presentation attack detection.IEEE Transactions on Information Forensics and Security, 18:3238–3250, 2023

    Aidan Boyd, Jeremy Speth, Lucas Parzianello, Kevin W Bowyer, and Adam Czajka. Comprehensive study in open- set iris presentation attack detection.IEEE Transactions on Information Forensics and Security, 18:3238–3250, 2023. 3

    work page 2023

  5. [5]

    An iris image synthesis method based on pca and super-resolution

    Jiali Cui, Yunhong Wang, JunZhou Huang, Tieniu Tan, and Zhenan Sun. An iris image synthesis method based on pca and super-resolution. InProceedings of the 17th In- ternational Conference on Pattern Recognition, 2004. ICPR 2004., volume 4, pages 471–474. IEEE, 2004. 2

    work page 2004

  6. [6]

    Domain-specific human-inspired binarized statisti- cal image features for iris recognition

    Adam Czajka, Daniel Moreira, Kevin Bowyer, and Patrick Flynn. Domain-specific human-inspired binarized statisti- cal image features for iris recognition. In2019 IEEE Win- ter Conference on Applications of Computer Vision (WACV), pages 959–967. IEEE, 2019. 4

    work page 2019

  7. [7]

    High confidence visual recognition of per- sons by a test of statistical independence

    John G Daugman. High confidence visual recognition of per- sons by a test of statistical independence. 15(11):1148–1161,

    work page

  8. [8]

    Improved training of wasserstein gans.Advances in neural information processing systems, 30, 2017

    Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans.Advances in neural information processing systems, 30, 2017. 3

    work page 2017

  9. [9]

    Information technology — Biometric data interchange formats — Part 6: Iris image data

    International Organization for Standardization. Information technology — Biometric data interchange formats — Part 6: Iris image data. ISO/IEC 19794-6:2011, 2011.https: //www.iso.org/standard/50869.html. 5

    work page 2011

  10. [10]

    Alias-free generative adversarial networks.Advances in neural infor- mation processing systems, 34:852–863, 2021

    Tero Karras, Miika Aittala, Samuli Laine, Erik H ¨ark¨onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks.Advances in neural infor- mation processing systems, 34:852–863, 2021. 7

    work page 2021

  11. [11]

    De- formirisnet: An identity-preserving model of iris texture de- formation

    Siamul Karim Khan, Patrick Tinsley, and Adam Czajka. De- formirisnet: An identity-preserving model of iris texture de- formation. InProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 900–908,

    work page

  12. [12]

    Synthetic iris presentation attack using idcgan

    Naman Kohli, Daksha Yadav, Mayank Vatsa, Richa Singh, and Afzel Noore. Synthetic iris presentation attack using idcgan. In2017 IEEE international joint conference on bio- metrics (IJCB), pages 674–680. IEEE, 2017. 3

    work page 2017

  13. [13]

    Conditional generative adversarial network-based data aug- mentation for enhancement of iris recognition accuracy

    Min Beom Lee, Yu Hwan Kim, and Kang Ryoung Park. Conditional generative adversarial network-based data aug- mentation for enhancement of iris recognition accuracy. IEEE Access, 7:122134–122152, 2019. 3

    work page 2019

  14. [14]

    Conditional wasserstein generative adversarial networks for rebalancing iris image datasets.IEICE TRANSACTIONS on Information and Sys- tems, 104(9):1450–1458, 2021

    Yung-Hui Li, Muhammad Saqlain Aslam, Latifa Nabila Harfiya, and Ching-Chun Chang. Conditional wasserstein generative adversarial networks for rebalancing iris image datasets.IEICE TRANSACTIONS on Information and Sys- tems, 104(9):1450–1458, 2021. 3

    work page 2021

  15. [15]

    Synthesis of iris images using markov random fields

    Sarvesh Makthal and Arun Ross. Synthesis of iris images using markov random fields. In2005 13th European Signal Processing Conference, pages 1–4. IEEE, 2005. 2

    work page 2005

  16. [16]

    Iris-GAN: Learning to Generate Realistic Iris Images Using Convolutional GAN

    Shervin Minaee and Amirali Abdolrashidi. Iris-gan: Learn- ing to generate realistic iris images using convolutional gan. arXiv preprint arXiv:1812.04822, 2018. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  17. [17]

    Rsgan: face swapping and editing using face and hair repre- sentation in latent spaces.arXiv preprint arXiv:1804.03447,

    Ryota Natsume, Tatsuya Yatagawa, and Shigeo Morishima. Rsgan: face swapping and editing using face and hair repre- sentation in latent spaces.arXiv preprint arXiv:1804.03447,

    work page arXiv

  18. [18]

    OSIRIS: An open source iris recognition software

    Nadia Othman, Bernadette Dorizzi, and Sonia Garcia- Salicetti. OSIRIS: An open source iris recognition software. Pattern Recognition Letters, 82:124–131, 2016. 4

    work page 2016

  19. [19]

    Synthetic iris im- age databases and identity leakage: Risks and mitigation strategies.arXiv preprint arXiv:2506.02626, 2025

    Ada Sawilska and Mateusz Trokielewicz. Synthetic iris im- age databases and identity leakage: Risks and mitigation strategies.arXiv preprint arXiv:2506.02626, 2025. 3

    work page arXiv 2025

  20. [20]

    Generating synthetic irises by feature agglomeration

    Samir Shah and Arun Ross. Generating synthetic irises by feature agglomeration. In2006 international conference on image processing, pages 317–320. IEEE, 2006. 2

    work page 2006

  21. [21]

    In- terpreting the latent space of gans for semantic face editing

    Yujun Shen, Jinjin Gu, Xiaoou Tang, and Bolei Zhou. In- terpreting the latent space of gans for semantic face editing. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 9243–9252, 2020. 3

    work page 2020

  22. [22]

    Latent traversals in generative models as potential flows

    Yue Song, T Anderson Keller, Nicu Sebe, and Max Welling. Latent traversals in generative models as potential flows. arXiv preprint arXiv:2304.12944, 2023. 3

    work page arXiv 2023

  23. [23]

    Warpedganspace: Finding non-linear rbf paths in gan latent space

    Christos Tzelepis, Georgios Tzimiropoulos, and Ioannis Pa- tras. Warpedganspace: Finding non-linear rbf paths in gan latent space. InProceedings of the IEEE/CVF international conference on computer vision, pages 6393–6402, 2021. 3

    work page 2021

  24. [24]

    Generating intra-and inter-class iris images by identity con- trast

    Chen Wang, Zhaofeng He, Caiyong Wang, and Qing Tian. Generating intra-and inter-class iris images by identity con- trast. In2022 IEEE International Joint Conference on Bio- metrics (IJCB), pages 1–7. IEEE, 2022. 1, 2, 3

    work page 2022

  25. [25]

    Iris synthesis: a reverse subdivision application

    Lakin Wecker, Faramarz Samavati, and Marina Gavrilova. Iris synthesis: a reverse subdivision application. InProceed- ings of the 3rd international conference on Computer graph- ics and interactive techniques in Australasia and South East Asia, pages 121–125, 2005. 2

    work page 2005

  26. [26]

    A multiresolution approach to iris synthesis.Computers & Graphics, 34(4):468–478, 2010

    Lakin Wecker, Faramarz Samavati, and Marina Gavrilova. A multiresolution approach to iris synthesis.Computers & Graphics, 34(4):468–478, 2010. 2

    work page 2010

  27. [27]

    Synthesis of large realistic iris databases using patch-based sampling

    Zhuoshi Wei, Tieniu Tan, and Zhenan Sun. Synthesis of large realistic iris databases using patch-based sampling. In2008 19th International Conference on Pattern Recognition, pages 1–4. IEEE, 2008. 2, 3

    work page 2008

  28. [28]

    Gan inversion: A survey

    Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 45(3):3121–3138, 2023. 4

    work page 2023

  29. [29]

    Synthesiz- ing iris images using rasgan with application in presentation attack detection

    Shivangi Yadav, Cunjian Chen, and Arun Ross. Synthesiz- ing iris images using rasgan with application in presentation attack detection. InProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition workshops, pages 0–0, 2019. 3

    work page 2019

  30. [30]

    Cit-gan: Cyclic image translation generative adversarial network with application in iris presentation attack detection

    Shivangi Yadav and Arun Ross. Cit-gan: Cyclic image translation generative adversarial network with application in iris presentation attack detection. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2412–2421, 2021. 3

    work page 2021

  31. [31]

    iwarpgan: Disentangling identity and style to generate synthetic iris images

    Shivangi Yadav and Arun Ross. iwarpgan: Disentangling identity and style to generate synthetic iris images. In2023 IEEE International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2023. 2, 3

    work page 2023

  32. [32]

    Synthesizing iris images us- ing generative adversarial networks: survey and comparative analysis.arXiv preprint arXiv:2404.17105, 2024

    Shivangi Yadav and Arun Ross. Synthesizing iris images us- ing generative adversarial networks: survey and comparative analysis.arXiv preprint arXiv:2404.17105, 2024. 3

    work page arXiv 2024

  33. [33]

    A model based, anatomy based method for synthesizing iris images

    Jinyu Zuo and Natalia A Schmid. A model based, anatomy based method for synthesizing iris images. InInternational Conference on Biometrics, pages 428–435. Springer, 2006. 2

    work page 2006

  34. [34]

    On genera- tion and analysis of synthetic iris images.IEEE Transactions on Information Forensics and Security, 2(1):77–90, 2007

    Jinyu Zuo, Natalia A Schmid, and Xiaohan Chen. On genera- tion and analysis of synthetic iris images.IEEE Transactions on Information Forensics and Security, 2(1):77–90, 2007. 2

    work page 2007