Pith. sign in

REVIEW 2 major objections 5 minor 62 references

Facial Demorphing from a Single Morph Using a Latent Conditional GAN

T0 review · 2 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 A single face morph can be split into its two constituent identities by a conditional GAN that works entirely in a pretrained autoencoder's latent space, even when the morph was made by an unseen technique.

desk verdict A real latent-space demorphing method with strong results on landmark morphs, but the agnosticism claim collapses on StyleGAN morphs and a suspicious SSIM=1.0 baseline row needs explanation. read the letter →

arxiv 2507.18566 v2 pith:6G44CTLG submitted 2025-07-24 cs.CV

Nitish Shukla , Arun Ross This is my paper
classification cs.CV
keywords facemorphingdemorphinglatentspaceconditionalGANbiometricforensicsmorphattackdetectionkurtosislossrecognition
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

A face morph combines two people's photos into one image that matches both identities, and existing demorphing methods either reproduce the morph itself or only work when training and testing use the same morphing software. This paper proposes a reference-free demorpher that first compresses the morph into a perceptually meaningful latent space with a frozen autoencoder, then trains a conditional GAN to decompose that latent representation into two separate face latents, which are decoded back into images. Trained exclusively on synthetic faces morphed with one landmark-based algorithm, the system is claimed to demorph real faces morphed with several unseen techniques and styles. If correct, a single trained demorpher could provide forensic evidence that identifies the two people behind a morph without knowing how the morph was made and without any reference image.

What carries the argument

The load-bearing mechanism is the frozen latent compression: a pretrained autoencoder maps each $512\times512$ face to a $64\times64\times4$ latent tensor, so identity-relevant semantics are kept while pixel-space noise, background, lighting, and morphing artifacts are suppressed. A conditional GAN then operates entirely in this latent space: the generator is a conditional UNet with the morph latent as condition, the discriminator scores triplets of latent encodings, and a kurtosis loss $L_{\mathrm{kurt}} = \sum_{j=1}^2 |\mathrm{Kurt}(o_j) - \mathrm{Kurt}(i_j)|$ acts as a separation prior that prevents the two outputs from collapsing into a copy of the morph. The decoder is used only at inference to turn the two output latents back into face images.

What would settle it

Generate a large set of morphs with a deep morphing architecture that operates in a very different feature domain, such as a neural implicit morph, hand the model one morph at a time, and check the two outputs against ground-truth identities with a face matcher; if restoration accuracy falls to the level where both outputs match the same constituent or the morph itself, the claimed technique-agnostic demorphing fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that demorphing can be done in the latent space of a pretrained image autoencoder rather than in raw pixel space, and that this makes the decomposition generalize across morphing techniques. The method, Latent Conditional GAN, encodes a single morph with the frozen Stable Diffusion encoder, uses a conditional UNet generator conditioned on that encoding to produce two constituent face encodings, and trains with an adversarial discriminator, an L1 reconstruction loss, and a kurtosis loss that aligns the higher-order statistics of the outputs with those of real faces. This combination is claimed to suppress the morph-replication failure mode, where both outputs collapse to the morph itself, and to yield high-fidelity faces that match the ground-truth identities. Under a unified protocol, the authors report restoration accuracy above 95 percent for landmark-based morphs and for MorDiff morphs, outperforming the three prior reference-free methods by large margins; StyleGAN-generated morphs remain difficult, with restoration accuracy around 12 to 50 percent depending on the face matcher.

Load-bearing premise

The load-bearing premise is that the frozen autoencoder's latent space keeps enough identity information to tell the two faces apart while throwing away the morphing technique's pixel-level artifacts; if identity and artifact are entangled in that space, the generator cannot separate them.

Editorial extensions

If this is right

  • A demorpher trained once on landmark-based morphs of synthetic faces can be applied directly to real morphs made with OpenCV, FaceMorpher, WebMorph, and MorDIFF without retraining.
  • Because no reference image is needed, the method can be used on a single suspect image to produce two candidate identities for forensic follow-up.
  • Demorphing in latent space rather than pixel space is what suppresses morph replication, so the two outputs are visually distinct from each other and from the input morph.
  • The near-perfect restoration accuracy on landmark-based and diffusion-based morphs indicates the decomposition is capturing identity, not just producing a face-like output.
  • StyleGAN morphs remain a hard case, with restoration accuracy dropping to about 12 to 50 percent, so the claimed generality is not yet uniform across all deep morphing techniques.

Reading between the lines

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

  • A natural extension is to treat the frozen autoencoder as a replaceable component: if a better identity-preserving compressor exists, the same conditional-GAN training could transfer to other morph families, including deep-generative ones that currently fail.
  • The kurtosis separation prior suggests a more general principle: aligning higher-order output statistics with real data can prevent collapse in any ill-posed decomposition task, not just faces.
  • The StyleGAN shortfall implies that the latent space is not fully invariant to generative-morph artifacts; a testable remedy would be to mix morphing techniques in training or to fine-tune the autoencoder on morph reconstructions.
  • Deployment would benefit from a confidence score or rejection rule, since the current pipeline always returns two faces, even on inputs where the decomposition is unreliable.
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

2 major / 5 minor

Summary. The paper proposes a reference-free facial demorphing method, Latent Conditional GAN, which encodes a morph with a frozen Stable Diffusion autoencoder and trains a conditional GAN in latent space to reconstruct the two constituent face images. Training uses synthetic faces morphed with OpenCV/dlib landmark-based morphing, while evaluation is performed on AMSL, FRLL-Morph (OpenCV, FaceMorpher, WebMorph, StyleGAN), MorDiff, and a small live-subject study. The authors compare against IPD, SDeMorph, and Face Demorphing under a common protocol and report large improvements on landmark-based and MorDiff morphs. The central claim is that operating in the frozen latent space makes the method agnostic to morphing technique and face style, so that it can demorph images from unseen morph algorithms without retraining.

Significance. If the central claim were correct, this would be a valuable advance: a single reference-free demorpher trained once on synthetic data and applied to unseen morphing algorithms and face styles. The paper has notable strengths, including a unified comparison of prior methods, a large synthetic training set, explicit identity-leakage checks, and a live-subject evaluation. However, the paper's own results on StyleGAN morphs directly contradict the advertised generalization, and an ablation baseline in Table 3 raises questions about how the absolute image-quality numbers are computed. As presented, the method generalizes well to landmark-based morphs and MorDiff, but not to at least one important deep-generative morphing technique, which is exactly the kind of unseen technique the abstract claims to handle.

major comments (2)
  1. [Abstract, Section 1, Table 1] The abstract and Section 1 claim that the method can demorph images created from unseen morph techniques and face styles and is agnostic to morphing technique. Table 1 shows that on FRLL-Morph StyleGAN morphs, Restoration Accuracy drops to 38.76% (ArcFace) and 50.39% (AdaFace) at 10% FMR, and to 2.12% and 4.52% at 0.1% FMR, while all landmark-based and MorDiff morphs reach near 100%. Since training uses only OpenCV/dlib landmark morphs (Section 5), StyleGAN is precisely an unseen technique, so this is a direct counterexample to the central claim rather than a minor degradation. The assertion in Section 4.1 that the frozen autoencoder discards morphing artifacts and standardizes the representation is not supported for StyleGAN morphs.
  2. [Table 3] In Table 3, the 'l1 image trivial' baseline reaches SSIM 1.0 on every dataset. If that baseline simply outputs the morph input, SSIM 1.0 can only be obtained if SSIM is computed against the morph rather than against the ground-truth constituent images. This looks like an evaluation artifact and makes the absolute SSIM/PSNR values in Table 2 difficult to interpret. Please clarify what this baseline is, which reference image is used for its SSIM, and why its SSIM is exactly 1.0 on all datasets.
minor comments (5)
  1. [Table 3 vs. Table 1] The StyleGAN Restoration Accuracy values are inconsistent across tables: Table 1 reports 38.76% for ArcFace and 50.39% for AdaFace at 10% FMR, while Table 3 reports 38.56/50.39 under an AdaFace/ArcFace heading without stating the FMR threshold. Please reconcile the numbers and specify the threshold used in Table 3.
  2. [Section 6.1] Restoration Accuracy is a primary evaluation metric, but it is never formally defined in the text. Please provide its definition so that the reader can interpret the reported percentages.
  3. [Table 2, Abstract, Section 9] The reported PSNR values are roughly 10–12 dB and SSIM values are roughly 0.5, which are low in absolute terms. The phrase 'high fidelity' used in the abstract and conclusion should be justified or tempered with respect to this metric scale.
  4. [Section 4.3] The UNet timestep is fixed to zero in all experiments. Since this removes a central conditioning mechanism of the UNet2DConditionModel, a brief justification for this design choice should be added.
  5. [Section 8] The live-subject study uses 17 subjects and 28 morphs. The reported Restoration Accuracy of 91–96% should be presented with an appropriate caveat about the small sample size.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the method is trained with supervised pixel/latent losses, and its evaluation metrics are standard and applied uniformly; the main risk is empirical generalization, not circularity.

full rationale

I walked the derivation chain. The demorpher is trained by minimizing Eq. (6), the sum of a conditional GAN loss, an L1 reconstruction loss in the frozen Stable Diffusion latent space, and a kurtosis regularizer, on synthetic OpenCV-generated morphs. The test-time outputs are decoded with E_dec and compared with ground-truth constituent images through ArcFace/AdaFace restoration accuracy and PSNR/SSIM/BW-IQA. None of these equations defines the predicted outputs in terms of a test-time fit, and no parameter is fitted to the FRLL/MorDiff test sets. The StyleGAN row of Table 1 (Rest. Acc. 38.76% at 10% FMR) contradicts the abstract's agnosticism claim, but that is an empirical generalization failure, not a circular reduction: the training objective does not presuppose the StyleGAN result. The self-citations ([52], [54], [55]) are used as baselines, a synthetic-data generation protocol, and an evaluation metric; the BW-IQA metric is applied uniformly to all compared methods and is not derived from the demorpher's outputs. I therefore find no step in which a 'prediction' or 'first-principles result' reduces by construction to its inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical or fictional entities are introduced. The method's additions are a new model architecture and loss terms; the load-bearing postulates are domain assumptions about the pretrained autoencoder's latent space and about cross-domain transfer.

free parameters (3)
  • lambda_1 (L1 loss weight) = 0.5
    Set to 0.5 in Section 4.2 without a sensitivity study; a hand-chosen hyperparameter.
  • lambda_2 (kurtosis loss weight) = 0.5
    Set to 0.5 in Section 4.2; the kurtosis term is central to the claimed morph-replication suppression, but no ablation varies its weight.
  • UNet timestep = 0
    Timestep fixed to zero during all training and inference to neutralize the denoising input (Section 4.3).
assumptions (3)
  • domain assumption The frozen Stable Diffusion autoencoder's latent space preserves semantic identity and suppresses morphing artifacts and background.
    The entire demorphing procedure operates in this latent space (Sections 4 and 4.1); if this premise fails, L1 and kurtosis losses in latent space cannot separate identities.
  • domain assumption Training on synthetic faces morphed with OpenCV/dlib transfers to real faces and to unseen morphing techniques.
    Section 5 states the training protocol; the StyleGAN results in Table 1 show this premise is only partially true.
  • ad hoc to paper Minimizing kurtosis differences between outputs and ground truths is a sufficient separation prior to avoid morph replication.
    Introduced in Section 4.2 with no derivation from morphing geometry; only empirically evaluated in Table 3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Facial Demorphing from a Single Morph Using a Latent Conditional GAN." pith.science (2026). https://pith.science/paper/6G44CTLG

@misc{pith2026250718566,
  author       = {Pith},
  title        = {Pith review of: Facial Demorphing from a Single Morph Using a Latent Conditional GAN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6G44CTLG}},
  note         = {Machine review of arXiv:2507.18566}
}
read the original abstract

A morph is created by combining two (or more) face images from two (or more) identities to create a composite image that is highly similar to all constituent identities, allowing the forged morph to be biometrically associated with more than one individual. Morph Attack Detection (MAD) can be used to detect a morph, but does not reveal the constituent images. Demorphing - the process of deducing the constituent images - is thus vital to provide additional evidence about a morph. Existing demorphing methods suffer from the morph replication problem, where the outputs tend to look very similar to the morph itself, or assume that train and test morphs are generated using the same morph technique. The proposed method overcomes these issues. The method decomposes a morph in latent space allowing it to demorph images created from unseen morph techniques and face styles. We train our method on morphs created from synthetic faces and test on morphs created from real faces using different morph techniques. Our method outperforms existing methods by a considerable margin and produces high fidelity demorphed face images.

Figures

Figures reproduced from arXiv: 2507.18566 by the authors.

Figure 1
Figure 1. Comparison with current state-of-the-art demorphing methods under a unified training protocol: Our method surpasses current [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proposed Demorphing Architecture: An encoder, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Scenarios in demorphing: Green arrows indicate a valid [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Reference-free demorphing: We test our method on the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Our method ensures that both outputs maintain the same [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: A comparison of the results based on different image separation priors: Zhang’s exclusion loss [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Testing on live subjects: Our method effectively separates the ground truth identities based on gender and facial features. All [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 60 canonical work pages

  1. [1]

    Article 6(4)(c) of the general data protection regula- tion

    The European Parliament and the Council of the European Union. Article 6(4)(c) of the general data protection regula- tion. 2016. 5

    work page 2016

  2. [2]

    Article 9(2)(a) of the general data protection regula- tion

    The European Parliament and the Council of the European Union. Article 9(2)(a) of the general data protection regula- tion. 2016. 5

    work page 2016

  3. [3]

    Article 37(1) of the general data protection regula- tion

    The European Parliament and the Council of the European Union. Article 37(1) of the general data protection regula- tion. 2016. 5

    work page 2016

  4. [4]

    Article 6(1)(a) of the general data protection regula- tion

    The European Parliament and the Council of the European Union. Article 6(1)(a) of the general data protection regula- tion. 2016. 5

    work page 2016

  5. [5]

    Article 9 of the general data protection regulation

    The European Parliament and the Council of the European Union. Article 9 of the general data protection regulation

    work page

  6. [6]

    The European Parliament and the Council of the European Union. Regulation (EU) 2016/679 of the european parlia- ment and of the council of 27 April 2016 on the protection of nat ural persons with regard to the processing of personal data and on the free movement of such data, and repeal- ing direc tive 95/46/ec (General Data Protection Regulation)

    work page 2016

  7. [7]

    Facial De-morphing: Extracting Component Faces from a Single Morph

    Sudipta Banerjee, Prateek Jaiswal, and Arun Ross. Facial De-morphing: Extracting Component Faces from a Single Morph. In Proceedings of IEEE International Joint Confer- ence on Biometrics, 2022. 1, 2, 3, 4, 6, 8

    work page 2022

  8. [8]

    Conditional identity disen- tanglement for differential face morph detection

    Sudipta Banerjee and Arun Ross. Conditional identity disen- tanglement for differential face morph detection. InProceed- ings of IEEE International Joint Conference on Biometrics (IJCB), pages 1–8, 2021. 1

    work page 2021

Show all 62 references
  1. [9]

    Feature Interaction-Based Face De- Morphing Factor Prediction for Restoring Accomplice’s Fa- cial Image

    Juan Cai, Qiangqiang Duan, Min Long, Le-Bing Zhang, and Xiangling Ding. Feature Interaction-Based Face De- Morphing Factor Prediction for Restoring Accomplice’s Fa- cial Image. Sensors, 24(17), 2024. 1

    2024

  2. [10]

    MorDIFF: Recognition Vulnerability and Attack Detectability of Face Morphing At- tacks Created by Diffusion Autoencoders

    Naser Damer, Meiling Fang, Patrick Siebke, Jan Niklas Kolf, Marco Huber, and Fadi Boutros. MorDIFF: Recognition Vulnerability and Attack Detectability of Face Morphing At- tacks Created by Diffusion Autoencoders. In Proceedings of 11th International Workshop on Biometrics and ...

    2023

  3. [11]

    On the generalization of detecting face morphing attacks as anomalies: Novelty vs

    Naser Damer, Jonas Henry Grebe, Steffen Zienert, Florian Kirchbuchner, and Arjan Kuijper. On the generalization of detecting face morphing attacks as anomalies: Novelty vs. outlier detection. In Proceedings of IEEE 10th International Conference on Biometrics Theory, Applicatio...

    2019

  4. [12]

    MorGAN: Recognition Vulnerability and Attack Detectability of Face Morphing Attacks Created by Generative Adversarial Network

    Naser Damer, Alexandra Mosegu ´ı Saladi ´e, Andreas Braun, and Arjan Kuijper. MorGAN: Recognition Vulnerability and Attack Detectability of Face Morphing Attacks Created by Generative Adversarial Network. In Proceedings of 2018 IEEE 9th International Conference on Biometrics T...

    2018

  5. [13]

    To detect or not to detect: The right faces to morph

    Naser Damer, Alexandra Mosegu ´ı Saladi ´e, Steffen Zienert, Yaza Wainakh, Philipp Terh¨orst, Florian Kirchbuchner, and Arjan Kuijper. To detect or not to detect: The right faces to morph. In Proceedings of International Conference on Biometrics (ICB), pages 1–8, 2019. 6

    2019

  6. [14]

    debruine/webmorph morphing software: Beta release 2, Jan

    Lisa DeBruine. debruine/webmorph morphing software: Beta release 2, Jan. 2018. 2, 5

    2018

  7. [15]

    Face Research Lab Lon- don (FRLL) Image Dataset

    Lisa DeBruine and Benedict Jones. Face Research Lab Lon- don (FRLL) Image Dataset. May 2017. 5

    2017

  8. [16]

    Arcface: Additive Angular Margin Loss for Deep Face Recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive Angular Margin Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 4690–4699, 2019. 5

    2019

  9. [17]

    Sigmoid- weighted linear units for neural network function approxi- mation in reinforcement learning, 2017

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid- weighted linear units for neural network function approxi- mation in reinforcement learning, 2017. 5

    2017

  10. [18]

    The magic passport

    Matteo Ferrara, Annalisa Franco, and Davide Maltoni. The magic passport. In IEEE International Joint Conference on Biometrics, 2014. 5

    2014

  11. [19]

    Face Demorphing

    Matteo Ferrara, Annalisa Franco, and Davide Maltoni. Face Demorphing. IEEE Transactions on Information Forensics and Security, 13(4):1008–1017, 2018. 1, 5

    2018

  12. [20]

    Face demorphing in the presence of facial appearance variations

    Matteo Ferrara, Annalisa Franco, and Davide Maltoni. Face demorphing in the presence of facial appearance variations. In Proceedings of European Signal Processing Conference (EUSIPCO), pages 2365–2369, 2018. 1

    2018

  13. [21]

    On the impact of alterations on face photo recog- nition accuracy

    Matteo Ferrara, Annalisa Franco, Davide Maltoni, and Yun- lian Sun. On the impact of alterations on face photo recog- nition accuracy. In Proceedings of International Conference on Image Analysis and Processing (ICIAP), 2013. 1

    2013

  14. [22]

    Accelerate: Training and inference at scale made simple, efficient and adaptable

    Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar, Marc Sun, and Benjamin Bossan. Accelerate: Training and inference at scale made simple, efficient and adaptable. https: //github.com/huggingface/accelerate , 2022. 5

    2022

  15. [23]

    Denoising Dif- fusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Dif- fusion Probabilistic Models. In Proceedings of Advances in Neural Information Processing Systems , volume 33, pages 6840–6851, 2020. 3

    2020

  16. [24]

    Deep metric learning using triplet network

    Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Proceedings of International Workshop on Similarity-Based Pattern Recognition, 2014. 7

    2014

  17. [25]

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversar- ial networks. In Proceedings of IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 5967– 5976, 2017. 4

    2017

  18. [26]

    Training generative ad- versarial networks with limited data

    Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative ad- versarial networks with limited data. In Proceedings of Ad- vances in Neural Information Processing Systems, 2020. 5

    2020

  19. [27]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396–4405, 2019. 6

    2019

  20. [28]

    Analyzing and Improv- ing the Image Quality of StyleGAN

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and Improv- ing the Image Quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5

    2020

  21. [29]

    Adaface: Quality Adaptive Margin for Face Recognition

    Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality Adaptive Margin for Face Recognition. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5

    2022

  22. [30]

    Davis E. King. Dlib-ml: A Machine Learning Toolkit. The Journal of Machine Learning Research, 10:1755–1758, Dec

  23. [31]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Represen- tations, (ICLR), 2015. 5

    2015

  24. [32]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Vari- ational Bayes. In Proceedings of International Conference on Learning Representations, (ICLR), 2014. 4

    2014

  25. [33]

    Adff: Adaptive de-morphing factor framework for restoring accomplice’s facial image

    Min Long, Jun Zhou, Le-Bing Zhang, Fei Peng, and Dengy- ong Zhang. Adff: Adaptive de-morphing factor framework for restoring accomplice’s facial image. IET Image Process- ing, 18(2):470–480, 2024. 1

    2024

  26. [34]

    Face morph using opencv — c++ / python

    Satya Mallick. Face morph using opencv — c++ / python. LearnOpenCV, 2016. 6

    2016

  27. [35]

    Conditional generative adversarial nets

    Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2

    2014 arXiv

  28. [36]

    Laws against morphing

    Matthias Monroy. Laws against morphing. Security Archi- tectures and the Police Collaboration in the EU, 2020. 1

    2020

  29. [37]

    Extended StirTrace Benchmarking of Biometric and Forensic Qualities of Mor- phed Face Images

    Tom Neubert, Andrey Makrushin, Mario Hildebrandt, Chris- tian Kraetzer, and Jana Dittmann. Extended StirTrace Benchmarking of Biometric and Forensic Qualities of Mor- phed Face Images. IET Biometrics, 7:325–332, 2018. 5

    2018

  30. [38]

    Face Recognition Vendor Test (FRVT) Part 4: MORPH - Performance of Automated Face Morph Detection

    Mei Ngan, Patrick Grother, Kayee Hanaoka, and Jason Kuo. Face Recognition Vendor Test (FRVT) Part 4: MORPH - Performance of Automated Face Morph Detection. NIST Interagency/Internal Report (NISTIR), National Institute of Standards and Technology, Gaithersburg, MD, 2020-03-06

    2020

  31. [39]

    FD-GAN: Face de-morphing generative adversarial network for restoring ac- complice’s facial image

    Fei Peng, Le-Bing Zhang, and Min Long. FD-GAN: Face de-morphing generative adversarial network for restoring ac- complice’s facial image. IEEE Access, 2019. 1, 3

    2019

  32. [40]

    Face morpher

    Alyssa Quek. Face morpher. 2, 5

  33. [41]

    Raghavendra, Kiran B

    R. Raghavendra, Kiran B. Raja, Sushma Venkatesh, and Christoph Busch. Transferable Deep-CNN Features for De- tecting Digital and Print-Scanned Morphed Face Images. In Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1822– ...

    2017

  34. [42]

    Bommanna Raja, Matteo Ferrara, Annalisa Franco, Luuk J

    K. Bommanna Raja, Matteo Ferrara, Annalisa Franco, Luuk J. Spreeuwers, Ilias Batskos, Florens de Wit, Marta Gomez-Barrero, Ulrich Scherhag, Daniel Fischer, Sushma Krupa Venkatesh, Jag Mohan Singh, Guoqiang Li, Lo¨ıc Bergeron, Sergey Isadskiy, Raghavendra Ramachandra, Christian...

    2020

  35. [43]

    Bommanna Raja, Matteo Ferrara, Annalisa Franco, Luuk J

    K. Bommanna Raja, Matteo Ferrara, Annalisa Franco, Luuk J. Spreeuwers, Ilias Batskos, Florens de Wit, Marta Gomez-Barrero, Ulrich Scherhag, Daniel Fischer, Sushma Krupa Venkatesh, Jag Mohan Singh, Guoqiang Li, Lo¨ıc Bergeron, Sergey Isadskiy, Raghavendra Ramachandra, Christian...

    2020

  36. [44]

    High-Resolution Image Synthesis with Latent Diffusion Models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models . In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674–10685, June 2022. 2

    2022

  37. [45]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, June 2022. 4, 5

    2022

  38. [46]

    Vulnerability Analysis of Face Morphing Attacks from Landmarks and Generative Adversarial Net- works

    Eklavya Sarkar, Pavel Korshunov, Laurent Colbois, and S´ebastien Marcel. Vulnerability Analysis of Face Morphing Attacks from Landmarks and Generative Adversarial Net- works. arXiv preprint arXiv:2012.05344, 2020. 6

    2012 arXiv

  39. [47]

    Face morph using opencv — c++ / python

    Mallick Satya. Face morph using opencv — c++ / python

  40. [48]

    Neural Implicit Morphing of Face Images

    Guilherme Schardong, Tiago Novello, Hallison Paz, Iurii Medvedev, Vin ´ıcius da Silva, Luiz Velho, and Nuno Gonc ¸alves. Neural Implicit Morphing of Face Images. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 7321–7330, Ju...

    2024

  41. [49]

    Detection of Face Mor- phing Attacks Based on PRNU Analysis

    Ulrich Scherhag, Luca Debiasi, Christian Rathgeb, Christoph Busch, and Andreas Uhl. Detection of Face Mor- phing Attacks Based on PRNU Analysis. In Proceedings of IEEE Transactions on Biometrics, Behavior, and Identity Science, 1(4):302–317, 2019. 5

    2019

  42. [50]

    Towards detection of morphed face images in electronic travel documents

    Ulrich Scherhag, Christian Rathgeb, and Christoph Busch. Towards detection of morphed face images in electronic travel documents. In Proceedings of IAPR International Workshop on Document Analysis Systems (DAS), pages 187– 192, 2018. 1

    2018

  43. [51]

    Deep Face Representations for Differen- tial Morphing Attack Detection

    Ulrich Scherhag, Christian Rathgeb, Johannes Merkle, and Christoph Busch. Deep Face Representations for Differen- tial Morphing Attack Detection. IEEE Transactions on In- formation Forensics and Security, 15:3625–3639, 2020. 6

    2020

  44. [52]

    SDeMorph: Towards Better Facial De- morphing from Single Morph

    Nitish Shukla. SDeMorph: Towards Better Facial De- morphing from Single Morph. In Proceedings of IEEE In- ternational Joint Conference on Biometrics (IJCB), 2023. 1, 2, 3, 4, 6, 8

    2023

  45. [53]

    dc-GAN: Dual-Conditioned GAN for Face Demorphing From a Single Morph

    Nitish Shukla and Arun Ross. dc-GAN: Dual-Conditioned GAN for Face Demorphing From a Single Morph. arXiv preprint arXiv:2411.14494, 2024. 2

    2024 arXiv

  46. [54]

    Facial Demorphing via Iden- tity Preserving Image Decomposition

    Nitish Shukla and Arun Ross. Facial Demorphing via Iden- tity Preserving Image Decomposition. In Proceedings of IEEE International Joint Conference on Biometrics (IJCB) ,

  47. [55]

    Metric for evaluating per- formance of reference-free demorphing methods

    Nitish Shukla and Arun Ross. Metric for evaluating per- formance of reference-free demorphing methods. In Pro- ceedings of IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2025. 5, 6

    2025

  48. [56]

    Lempitsky

    Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast styl- ization. ArXiv, abs/1607.08022, 2016. 5

    2016 arXiv

  49. [57]

    https : / / www

    V on Raphael Thelen und Judith Horchert. https : / / www . spiegel . de / netzwelt / netzp - olitik / biometrie - im - reisepass - peng - kollektiv - schmuggelt - fotomontage - in - ausweis-a-1229418.html, 2018. 1

    2018

  50. [58]

    Group normalization

    Yuxin Wu and Kaiming He. Group normalization. In Com- puter Vision – ECCV 2018: 15th European Conference, Mu- nich, Germany, September 8-14, 2018, Proceedings, Part XIII, page 3–19, Berlin, Heidelberg, 2018. Springer-Verlag. 5

    2018

  51. [59]

    MIP- GAN—generating strong and high quality morphing attacks using identity prior driven gan

    Haoyu Zhang, Sushma Venkatesh, Raghavendra Ramachan- dra, Kiran Raja, Naser Damer, and Christoph Busch. MIP- GAN—generating strong and high quality morphing attacks using identity prior driven gan. IEEE Transactions on Bio- metrics, Behavior, and Identity Science, 3(3), 2021. 2

    2021

  52. [60]

    Zhang, Z

    K. Zhang, Z. Zhang, Z. Li, and Y . Qiao. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks. IEEE Signal Processing Letters , 23(10):1499– 1503, 2016. 6

    2016

  53. [61]

    Single Image Re- flection Separation with Perceptual Losses

    Xuaner Zhang, Ren Ng, and Qifeng Chen. Single Image Re- flection Separation with Perceptual Losses . In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4786–4794, June 2018. 7

    2018

  54. [62]

    Deep Adversarial Decomposition: A Unified Framework for Separating Superimposed Images

    Zhengxia Zou, Sen Lei, Tianyang Shi, Zhenwei Shi, and Jieping Ye. Deep Adversarial Decomposition: A Unified Framework for Separating Superimposed Images . In Pro- ceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12803–12813, 2020. 3, 4, 7

    2020

Pith tools

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