Pith. sign in

REVIEW 4 major objections 5 minor 69 references

WaFusion: A Wavelet-Enhanced Diffusion Framework for Face Morph Generation

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 WaFusion argues that a face morph can be generated by morphing only the low-frequency wavelet sub-band with a diffusion autoencoder and averaging the high-frequency sub-bands, yielding 512x512 morphs that evade FaceNet more often than…

desk verdict WaFusion is a sensible wavelet/diffusion efficiency hack for face morphing, but its headline claim over full-image diffusion morphs is resolution-confounded and the load-bearing ablation uses perceptual metrics instead of identity or attack metrics. read the letter →

arxiv 2507.12493 v1 pith:CJIYADTR submitted 2025-07-15 cs.GR

Seyed Rasoul Hosseini , Omid Ahmadieh , Jeremy Dawson , Nasser Nasrabadi This is my paper
classification cs.GR
keywords facemorphingwaveletdecompositiondiffusionautoencoderattackdetectionNetverificationHaarbiometricsecurity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes WaFusion, a face-morphing method that claims a convincing morph of two faces can be produced by applying a diffusion autoencoder to only the low-frequency (LL) wavelet sub-band of the two aligned images, while the high-frequency sub-bands are combined by simple averaging. The claim is that the LL sub-band holds the structural essence of a face, so morphing it and reconstructing with the inverse wavelet transform yields 512x512 morphs that are more challenging to a FaceNet verifier than full-image diffusion morphs, at roughly a quarter of the generative computation. On the FERET, FRGC, FRLL, and WVU Twin datasets, the paper reports WaFusion morphs are more challenging to FaceNet than landmark-based and full-image diffusion morphs, and on FERET, FRGC, and WVU Twin they also beat the GAN baseline; on FRLL, StyleGAN edges it out on some metrics. If this is right, morphing attacks can be made both cheaper and harder to detect, giving morphing-attack detection systems a stronger stress test.

What carries the argument

The central mechanism is the wavelet sub-band division of labor: a single-level Haar discrete wavelet transform splits each aligned 512x512 face into a 256x256 low-frequency approximation (LL) and three high-frequency detail sub-bands (LH, HL, HH). The generative workload is confined to the LL sub-band, which the paper calls the structural essence of the image; the diffusion autoencoder's dual encoders separate semantic content (linearly interpolated) from stochastic texture (spherically interpolated), and decode a morphed LL. The high-frequency sub-bands are averaged rather than generated, and the inverse wavelet transform recombines the morphed structure with the averaged detail. The mechanism is carried by the assertion that the LL sub-band contains enough identity information that averaging the high-frequency bands is perceptually and biometrically neutral, an assertion the paper supports with an ablation (Table 2) showing LPIPS and SSIM parity between LL-only and all-sub-band morphing on the WVU Twin dataset.

What would settle it

An experiment that would settle it: reconstruct a face from only the averaged high-frequency sub-bands by setting the morphed LL sub-band to zero before the inverse wavelet transform, then run FaceNet verification between that reconstruction and each source identity. If either source is identified at above-chance rate from high-frequency bands alone, those bands carry identity information that the averaging step may be discarding, and the claim that the LL sub-band holds the structural essence would be weakened.

Watch

Extended reading notes

Core claim

WaFusion's central discovery is that full-image generative morphing is wasteful: the generative model only needs the low-frequency structure, and the high-frequency detail can be handled non-generatively. Each aligned face is decomposed with a single-level Haar wavelet transform into four sub-bands (LL, LH, HL, HH). The LL sub-bands from the two subjects are fed into a diffusion autoencoder whose semantic-encoder streams are blended linearly and whose stochastic-encoder streams are blended by spherical linear interpolation; the decoded result is the morphed LL sub-band. The high-frequency sub-bands are averaged between the two subjects, and the inverse wavelet transform fuses the morphed LL with the averaged detail bands into a 512x512 morph. Across FERET, FRGC, FRLL, and WVU Twin, the authors report that WaFusion is the most attack-favorable on FERET, FRGC, and WVU Twin on the FaceNet metrics, and is competitive with StyleGAN on FRLL, while beating the landmark-based and full-diffusion baselines; an ablation shows that LL-only morphing matches all-sub-band morphing on LPIPS and SSIM while using about a quarter of the computation.

Load-bearing premise

The load-bearing premise is that the blurry low-frequency wavelet sub-band carries enough of what makes a face identifiable that morphing only that part and averaging the fine-detail sub-bands still yields a realistic morph of both people; if identity-critical details live in the high-frequency bands, such as the iris, skin texture, or scars, the averaging step could wash them out or introduce artifacts.

Editorial extensions

If this is right

  • WaFusion produces 512x512 morphs with the diffusion model running at 256x256 resolution, cutting the generative computation roughly fourfold relative to morphing all four sub-bands.
  • On FERET, FRGC, and WVU Twin, WaFusion morphs yield the lowest AUC and highest EER for FaceNet among the compared methods, meaning those verifiers are more often fooled; on FRLL, StyleGAN is competitive on some metrics.
  • The high-frequency averaging step is perceptually neutral on the WVU Twin set, with LPIPS and SSIM scores comparable to morphing all sub-bands.
  • Because the same frozen diffusion autoencoder serves across FERET, FRGC, FRLL, and WVU Twin, the pipeline transfers across datasets with differing resolution, pose, and subject similarity without retraining the generative model.
  • The 512x512 output resolution is achieved without raising the diffusion model's operating resolution, so the efficiency gain and the quality gain come from the same design choice.

Reading between the lines

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

  • The paper does not measure how much identity information survives in the high-frequency sub-bands; a direct test would reconstruct an image from only the averaged high-frequency bands and check whether FaceNet can still identify either subject, which would reveal whether the averaging step is truly neutral.
  • The same LL-morph-plus-average-detail recipe could apply to other biometric modalities, such as iris or fingerprint images with a low-frequency structural core, but the paper only demonstrates it for faces.
  • A natural testable extension is to vary the wavelet basis and decomposition level (e.g., Daubechies or multi-level Haar), which the authors list as future work; one would expect a trade-off curve between morph quality, attack success, and computational cost that the single-level Haar results do not yet provide.
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 paper proposes WaFusion, a face-morphing framework that combines a single-level Haar discrete wavelet transform with diffusion autoencoders. The two input faces are decomposed into four sub-bands; the LL sub-bands are morphed through semantic and stochastic latent interpolation in a pre-trained diffusion autoencoder, the high-frequency sub-bands (LH, HL, HH) are averaged, and the inverse wavelet transform reconstructs a 512×512 morph. The authors evaluate WaFusion against OpenCV, FaceMorpher, StyleGAN, WebMorpher, and a 256×256 diffusion autoencoder baseline on FERET, FRGC, FRLL, and WVU Twin, reporting AUC, APCER, BPCER, and EER with FaceNet as the verifier, plus an LL-only ablation using LPIPS/SSIM. The central claim is that wavelet-domain processing of the LL sub-band yields high-resolution, realistic morphs that are more challenging to face recognition than full-image diffusion morphs, at lower computational cost.

Significance. The idea of restricting the generative diffusion process to the LL wavelet sub-band while cheaply averaging the high-frequency sub-bands is a plausible route to efficient high-resolution morph generation, and the paper applies it to a practically important biometric problem. The evaluation has notable strengths: four datasets with differing similarity structure, a large number of generated morphs, standard ISO-style metrics, and an external FaceNet verifier rather than a self-reported quality score. If the central claim is fully supported after revision, the method would be a useful benchmark for morphing-attack research. However, the current experiments do not yet establish the claimed superiority over full-image diffusion morphing at matched resolution, and the ablation evidence for the load-bearing high-frequency-averaging assumption is indirect.

major comments (4)
  1. [§4.4, Table 1] The central comparative claim—that WaFusion produces 512×512 morphs more challenging to FaceNet than full-image diffusion morphs at lower computational cost—is not tested against a matched full-image diffusion baseline. The only diffusion baseline is described in §4.4 as running at 256×256, whereas WaFusion diffuses only the 256×256 LL sub-band and reconstructs 512×512 via the inverse wavelet transform. DiffMorpher [67] and Fast-DiM [3], which are cited in §2, are not evaluated. Because the comparison conflates output resolution with wavelet-domain processing, the uniqueness claim in §3.3 is not supported by the experiments as reported.
  2. [§3.3, Table 2] The load-bearing premise that high-frequency sub-bands can be safely averaged rests on an ablation that reports only LPIPS and SSIM on the WVU Twin dataset, with no identity or attack metrics. These perceptual-similarity scores do not establish that the LL-only morph remains matched to both source identities (for example, similar FaceNet cosine distances to both subjects), nor that it preserves the attack strength measured in Table 1. The ablation should report identity-similarity scores to both subjects and at least one attack metric (such as AUC or EER) for all-sub-band versus LL-only morphs.
  3. [§4.3, §4.4] The evaluation protocol for APCER, BPCER, and EER is underspecified. The paper does not describe how the morph and bona fide score distributions are constructed, which enrollment image is used, whether differential or single-image morphing-attack detection is assumed, how the thresholds at 5%, 10%, and 30% are chosen, or how many score comparisons underlie each cell in Table 1. Without this information, the results in Table 1 are not reproducible, and the meaning of the arrows in the table header is ambiguous.
  4. [§4.4, Table 1] All reported metrics are point estimates with no error bars, confidence intervals, or significance tests. On datasets with 529–2,971 morphs, differences of about one percentage point (e.g., FERET EER 51.856 for WaFusion versus 50.850 for the diffusion baseline) cannot be distinguished from sampling noise. The claim of superiority over baselines across datasets requires bootstrap estimates, confidence intervals, or statistical tests.
minor comments (5)
  1. [§1, first paragraph] 'Owning to' should be 'Owing to'.
  2. [§3.2, Eq. (6)] The displayed formula for q(x_{t-1} | x_t, x_0) is missing parentheses around (x_t - sqrt(alpha_t) x_0)/sqrt(1 - alpha_t); please rewrite it with clear notation and state the variance of the distribution.
  3. [§4.2] Please specify the exact pre-trained diffusion autoencoder checkpoint and the training data used, and state the number of DDIM sampling steps used for the diffusion baseline as well as for WaFusion.
  4. [Table 1, Figures 3 and 5] The table header 'EER (%)↑5% 10% 30%' is not readable; clarify whether EER is a single number and what thresholds the APCER and BPCER columns refer to, and add axis labels and a legend to the ROC and APCER–BPCER curves.
  5. [Figure 4 caption] Please identify what 'Diffusion' refers to (presumably Diffusion Autoencoders [38]) and state its resolution explicitly in the caption.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the wavelet-diffusion pipeline is self-contained and evaluated against an external FaceNet verifier; only a minor non-load-bearing self-citation is present.

full rationale

WaFusion's derivation chain is self-contained. The method decomposes two aligned 512x512 input images with a single-level Haar transform, runs the 256x256 LL sub-bands through a pre-trained diffusion autoencoder (Preechakul et al. [38]), interpolates semantic codes linearly and stochastic codes via spherical interpolation (Eq. 7), averages the LH/HL/HH sub-bands, and reconstructs the morph with the inverse wavelet transform. No parameter or constant in this pipeline is fitted to the reported AUC/APCER/BPCER/EER values; the evaluation uses an external FaceNet verifier pre-trained on VGGFace2. The load-bearing premise that the LL sub-band captures structural identity while the high-frequency sub-bands can be safely averaged is a stated design assumption supported only by an LPIPS/SSIM ablation (Table 2); that is an evidentiary weakness, not circularity, because the ablation compares two variants of the same framework and the attack metrics are not used to define the method. The citation to O'Haire et al. [33], from the same group, for wavelet-based morphing is not load-bearing: the wavelet equations are standard (Eq. 1) and the diffusion autoencoder is external. The comparison against a 256x256 full-image diffusion baseline and the absence of identity-similarity metrics in the ablation are experimental fairness and validity concerns, not circularity. Therefore no step in the paper reduces, by construction or self-citation, to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical entities. It relies on standard wavelet math (Haar, perfect reconstruction), a pre-trained diffusion autoencoder from Preechakul et al., and hand-chosen hyperparameters such as blending ratio and DDIM steps. The most consequential assumptions are the sufficiency of the LL sub-band for identity blending and the validity of FaceNet-based evaluation.

free parameters (2)
  • Blending ratio gamma = not reported, presumed 0.5
    Used in image-space and latent-space interpolation (Eq. 7 in Section 3.3); the paper does not state its value or tuning. It controls the identity blend and could affect attack strength.
  • DDIM sampling steps = 100
    Chosen by hand in Section 4.2; affects output fidelity and computation, though not a fitted parameter.
assumptions (4)
  • standard math The Haar wavelet transform with the inverse wavelet transform achieves lossless reconstruction of the original image.
    Invoked in Section 3.1 to justify fusing morphed LL and averaged high-frequency sub-bands without loss.
  • domain assumption The pre-trained diffusion autoencoder (Preechakul et al.) provides a semantically meaningful, decodable latent space for face images.
    Used in Section 3.3 to generate the morphed LL sub-band; the paper does not train or fine-tune this model.
  • domain assumption The LL wavelet sub-band contains sufficient identity and structural information that high-frequency sub-bands can be averaged without losing morph quality.
    Core design assumption in Section 3.3, supported only by an ablation on one dataset (Table 2).
  • domain assumption FaceNet, pre-trained on VGGFace2, yields embeddings appropriate for measuring morph attack success via APCER, BPCER, and EER.
    Used in Section 4.2 for all reported metrics; no justification of the verifier's suitability for this protocol is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WaFusion: A Wavelet-Enhanced Diffusion Framework for Face Morph Generation." pith.science (2026). https://pith.science/paper/CJIYADTR

@misc{pith2026250712493,
  author       = {Pith},
  title        = {Pith review of: WaFusion: A Wavelet-Enhanced Diffusion Framework for Face Morph Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CJIYADTR}},
  note         = {Machine review of arXiv:2507.12493}
}
read the original abstract

Biometric face morphing poses a critical challenge to identity verification systems, undermining their security and robustness. To address this issue, we propose WaFusion, a novel framework combining wavelet decomposition and diffusion models to generate high-quality, realistic morphed face images efficiently. WaFusion leverages the structural details captured by wavelet transforms and the generative capabilities of diffusion models, producing face morphs with minimal artifacts. Experiments conducted on FERET, FRGC, FRLL, and WVU Twin datasets demonstrate WaFusion's superiority over state-of-the-art methods, producing high-resolution morphs with fewer artifacts. Our framework excels across key biometric metrics, including the Attack Presentation Classification Error Rate (APCER), Bona Fide Presentation Classification Error Rate (BPCER), and Equal Error Rate (EER). This work sets a new benchmark in biometric morph generation, offering a cutting-edge and efficient solution to enhance biometric security systems.

Figures

Figures reproduced from arXiv: 2507.12493 by the authors.

Figure 1
Figure 1. Overview of the WaFusion framework for face morphing. The input images, Subject 1 and Subject 2, are first aligned and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture for diffusion-based morph generation using Diffusion Autoencoders. LL sub-bands (red dashed lines) from Subject [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. APCER-BPCER curves for FaceNet verifier across (a) FERET, (b) FRGC, (c) FRLL, and (d) WVU Twin datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visual comparison of morph images generated by [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: ROC curves for FaceNet verifier across (a) FERET, (b) FRGC, (c) FRLL, and (d) WVU Twin datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Morph images generated by different methods on the FRGC and FERET datasets, showing bona fide images (blue boxes) and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 58 canonical work pages

  1. [67]

    Zhang, Y

    K. Zhang, Y . Zhou, X. Xu, B. Dai, and X. Pan. Diffmor- pher: Unleashing the capability of diffusion models for im- age morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7912– 7921, 2024

    work page 2024

  2. [3]

    Z. W. Blasingame and C. Liu. Fast-dim: Towards fast diffu- sion morphs. IEEE Security & Privacy, 2024

    work page 2024

  3. [1]

    Ashwini, D

    K. Ashwini, D. Nagajyothi, C. Ramakrishna, and V . Jyothi. Generation of high-quality realistic faces with StyleGAN. In 2023 4th IEEE Global Conference for Advancement in Tech- nology (GCAT), pages 1–7. IEEE, 2023

    work page 2023

  4. [2]

    Blasingame and C

    Z. Blasingame and C. Liu. Leveraging diffusion for strong and high quality face morphing attacks. IEEE Transactions on Biometrics, Behavior, and Identity Science, 2024

    work page 2024

  5. [4]

    K. W. Bowyer. Face recognition technology: security versus privacy.IEEE Technology and society magazine, 23(1):9–19, 2004

    work page 2004

  6. [5]

    Q. Cao, L. Shen, W. Xie, O. M. Parkhi, and A. Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on auto- matic face & gesture recognition (FG 2018) , pages 67–74. IEEE, 2018

    2018

  7. [6]

    T. F. Cootes, G. J. Edwards, and C. J. Taylor. Active appear- ance models. IEEE Transactions on pattern analysis and machine intelligence, 23(6):681–685, 2001

    work page 2001

  8. [7]

    Damer, K

    N. Damer, K. Raja, M. S ¨ußmilch, S. Venkatesh, F. Boutros, M. Fang, F. Kirchbuchner, R. Ramachandra, and A. Kuijper. Regenmorph: Visibly realistic GAN generated face morph- ing attacks by attack re-generation. In Advances in Visual Computing: 16th International Symposium, ISVC 2021, Vir- tual Event, October 4-6, 2021, Proceedings, Part I , pages 251–264....

    work page 2021

Show all 69 references
  1. [8]

    Damer, A

    N. Damer, A. M. Saladie, A. Braun, and A. Kuijper. Mor- GAN: Recognition vulnerability and attack detectability of face morphing attacks created by generative adversarial net- work. In 2018 IEEE 9th international conference on biomet- rics theory, applications and systems (BTA...

    2018

  2. [9]

    Daubechies

    I. Daubechies. Ten lectures on wavelets. SIAM, 1992

    1992

  3. [10]

    DeBruine

    L. DeBruine. debruine/webmorph: Beta release 2. Zenodo https://doi. org/10, 5281, 2018

    2018

  4. [11]

    DeBruine and B

    L. DeBruine and B. Jones. Face research lab london set. Psychol. Methodol. Des. Anal, 3, 2017

    2017

  5. [12]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat GANs on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  6. [13]

    Ferrara, A

    M. Ferrara, A. Franco, and D. Maltoni. The magic passport. IEEE International Joint Conference on Biometrics , pages 1–7, 2014

    2014

  7. [14]

    Ferrara, A

    M. Ferrara, A. Franco, and D. Maltoni. Decoupling texture blending and shape warping in face morphing. In2019 inter- national conference of the biometrics special interest group (BIOSIG), pages 1–5. IEEE, 2019

    2019

  8. [15]

    I. O. for Standardization. ISO/IEC 30107-3:2023: Informa- tion technology – Biometric presentation attack detection. Technical report, ISO, 2023

    2023

  9. [16]

    S. R. Godage, F. Løv ˚asdal, S. Venkatesh, K. Raja, R. Ra- machandra, and C. Busch. Analyzing human observer ability in morphing attack detection—where do we stand? IEEE Transactions on Technology and Society , 4(2):125– 145, 2022

    2022

  10. [17]

    Gomez-Barrero, K

    M. Gomez-Barrero, K. B. Raja, C. Rathgeb, A. F. Sequeira, M. Todisco, L. Colbois, and S. Marcel. On the detection of morphing attacks generated by GANs. 2022 Interna- tional Conference of the Biometrics Special Interest Group (BIOSIG), pages 1–5, 2022

    2022

  11. [18]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Gen- erative adversarial networks. Communications of the ACM , 63(11):139–144, 2020

    2020

  12. [19]

    Hamza, S

    M. Hamza, S. Tehsin, M. Humayun, M. F. Almufareh, and M. Alfayad. A comprehensive review of face morph genera- tion and detection of fraudulent identities. Applied Sciences, 12(24):12545, 2022

    2022

  13. [20]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion proba- bilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  14. [21]

    Huang, J

    Y . Huang, J. Huang, J. Liu, Y . Dong, J. Lv, and S. Chen. Wavedm: Wavelet-based diffusion models for image restora- tion. arXiv preprint arXiv:2305.13819, 2023

    2023 arXiv

  15. [22]

    D. ICAO. 9303-machine readable travel documents-part 9: Deployment of biometric identification and electronic stor- age of data in emrtds. International Civil Aviation Organi- zation (ICAO), 123, 2015

    2015

  16. [23]

    Ivanovska and V

    M. Ivanovska and V . ˇStruc. Face morphing attack detec- tion with denoising diffusion probabilistic models. In 2023 11th International Workshop on Biometrics and Forensics (IWBF), pages 1–6. IEEE, 2023

    2023

  17. [24]

    Karras, S

    T. Karras, S. Laine, and T. Aila. A style-based generator ar- chitecture for generative adversarial networks. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019

    2019

  18. [25]

    Kessler, K

    R. Kessler, K. Raja, J. Tapia, and C. Busch. Towards min- imizing efforts for morphing attacks—deep embeddings for morphing pair selection and improved morphing attack de- tection. Plos one, 19(5):e0304610, 2024

    2024

  19. [26]

    J. Kim, C. Oh, H. Do, S. Kim, and K. Sohn. Diffusion- driven gan inversion for multi-modal face image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10403–10412, 2024

    2024

  20. [27]

    G. R. Lee, R. Gommers, F. Waselewski, K. Wohlfahrt, and A. O’Leary. Pywavelets: A python package for wavelet anal- ysis. Journal of Open Source Software, 4(36):1237, 2019

    2019

  21. [28]

    Makrushin, D

    A. Makrushin, D. Siegel, and J. Dittmann. Simulation of border control in an ongoing web-based experiment for esti- mating morphing detection performance of humans. In Pro- ceedings of the 2020 ACM Workshop on Information Hiding and Multimedia Security, pages 91–96, 2020

    2020

  22. [29]

    S. Mallat. A theory for multiresolution signal decomposi- tion: The wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell., 11:674–693, 1989

    1989

  23. [30]

    S. Mallick. Face morph using opencv—c++/python. LearnOpenCV, 1(1), 2016

    2016

  24. [31]

    Neubert, A

    T. Neubert, A. Makrushin, M. Hildebrandt, C. Kraetzer, and J. Dittmann. Extended stirtrace benchmarking of biometric and forensic qualities of morphed face images. Iet Biomet- rics, 7(4):325–332, 2018

    2018

  25. [32]

    A. Q. Nichol and P. Dhariwal. Improved denoising diffu- sion probabilistic models. In International conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  26. [33]

    O’Haire, S

    K. O’Haire, S. Soleymani, B. Chaudhary, P. Aghdaie, J. M. Dawson, and N. M. Nasrabadi. Adversarially perturbed wavelet-based morphed face generation. 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 01–05, 2021

    2021

  27. [34]

    O’Haire, S

    K. O’Haire, S. Soleymani, B. Chaudhary, J. Dawson, and N. M. Nasrabadi. Identical twins face morph database gener- ation. In 2022 IEEE International Joint Conference on Bio- metrics (IJCB), pages 1–9. IEEE, 2022

    2022

  28. [35]

    P. J. Phillips, P. J. Flynn, T. Scruggs, K. W. Bowyer, J. Chang, K. Hoffman, J. Marques, J. Min, and W. Worek. Overview of the face recognition grand challenge. In2005 IEEE computer society conference on computer vision and pattern recogni- tion (CVPR’05), volume 1, pages 947–...

    2005

  29. [36]

    P. J. Phillips, H. Wechsler, J. Huang, and P. J. Rauss. The feret database and evaluation procedure for face-recognition algorithms. Image and vision computing , 16(5):295–306, 1998

    1998

  30. [37]

    Phung, Q

    H. Phung, Q. Dao, and A. Tran. Wavelet diffusion models are fast and scalable image generators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10199–10208, 2023

    2023

  31. [38]

    Preechakul, N

    K. Preechakul, N. Chatthee, S. Wizadwongsa, and S. Suwa- janakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10609–10619, 2021

    2022

  32. [39]

    Price, S

    S. Price, S. Soleymani, and N. M. Nasrabadi. Landmark en- forcement and style manipulation for generative morphing. In 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2022

    2022

  33. [40]

    A. Quek. Facemorpher, 2019

    2019

  34. [41]

    Raghavendra, K

    R. Raghavendra, K. B. Raja, and C. Busch. Detecting mor- phed face images. 2016 IEEE 8th International Conference on Biometrics Theory, Applications and Systems (BTAS) , pages 1–7, 2016

    2016

  35. [42]

    K. Raja, M. Ferrara, A. Franco, L. Spreeuwers, I. Batskos, F. De Wit, M. Gomez-Barrero, U. Scherhag, D. Fischer, S. K. Venkatesh, et al. Morphing attack detection-database, eval- uation platform, and benchmarking. IEEE transactions on information forensics and security, 16:433...

    2020

  36. [43]

    Ramachandra and C

    R. Ramachandra and C. Busch. Presentation attack detec- tion methods for face recognition systems: A comprehensive survey. ACM Computing Surveys (CSUR), 50(1):1–37, 2017

    2017

  37. [44]

    Ramachandra, S

    R. Ramachandra, S. Venkatesh, N. Damer, N. Vetrekar, and R. S. Gad. Multispectral imaging for differential face morph- ing attack detection: A preliminary study. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, pages 6185–6193, 2024

    2024

  38. [45]

    D. J. Robertson, R. S. Kramer, and A. M. Burton. Fraudulent ID using face morphs: Experiments on human and automatic recognition. PLoS One, 12(3):e0173319, 2017

    2017

  39. [46]

    Roich, R

    D. Roich, R. Mokady, A. H. Bermano, and D. Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on graphics (TOG), 42(1):1–13, 2022

    2022

  40. [47]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer. High-resolution image synthesis with latent diffu- sion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684– 10695, 2022

    2022

  41. [48]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convolu- tional networks for biomedical image segmentation. InMed- ical image computing and computer-assisted intervention– MICCAI 2015: 18th international conference, Munich, Ger- many, October 5-9, 2015, proceedings, part III 1...

    2015

  42. [49]

    M. S. E. Saadabadi, S. R. Malakshan, S. R. Hosseini, and N. M. Nasrabadi. Boosting unconstrained face recognition with targeted style adversary. In 2024 IEEE International Joint Conference on Biometrics (IJCB) , pages 1–11. IEEE, 2024

    2024

  43. [50]

    Sarkar, P

    E. Sarkar, P. Korshunov, L. Colbois, and S. Marcel. Vul- nerability analysis of face morphing attacks from land- marks and generative adversarial networks. arXiv preprint arXiv:2012.05344, 2020

    2012 arXiv

  44. [51]

    Sarkar, P

    E. Sarkar, P. Korshunov, L. Colbois, and S. Marcel. Are GAN-based morphs threatening face recognition? In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2959–2963. IEEE, 2022

    2022

  45. [52]

    Scherhag, L

    U. Scherhag, L. Debiasi, C. Rathgeb, C. Busch, and A. Uhl. Detection of face morphing attacks based on PRNU analy- sis. IEEE Transactions on Biometrics, Behavior, and Identity Science, 1(4):302–317, 2019

    2019

  46. [53]

    Scherhag, A

    U. Scherhag, A. Nautsch, C. Rathgeb, M. Gomez-Barrero, R. N. Veldhuis, L. Spreeuwers, M. Schils, D. Maltoni, P. Grother, S. Marcel, et al. Biometric systems under mor- phing attacks: Assessment of morphing techniques and vul- nerability reporting. In 2017 International Confere...

    2017

  47. [54]

    Scherhag, C

    U. Scherhag, C. Rathgeb, J. Merkle, and C. Busch. Deep face representations for differential morphing attack detec- tion. IEEE transactions on information forensics and secu- rity, 15:3625–3639, 2020

    2020

  48. [55]

    Schroff, D

    F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815–823, 2015

    2015

  49. [56]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilib- rium thermodynamics. In International conference on ma- chine learning, pages 2256–2265. PMLR, 2015

    2015

  50. [57]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion im- plicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  51. [58]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 1–9, 2015

    2015

  52. [59]

    Venkatesh, R

    S. Venkatesh, R. Ramachandra, K. Raja, and C. Busch. Face morphing attack generation and detection: A comprehen- sive survey. IEEE transactions on technology and society , 2(3):128–145, 2021

    2021

  53. [60]

    Venkatesh, H

    S. Venkatesh, H. Zhang, R. Ramachandra, K. Raja, N. Damer, and C. Busch. Can GAN generated morphs threaten face recognition systems equally as landmark based morphs?-vulnerability and detection. In 2020 8th Interna- tional Workshop on Biometrics and Forensics (IWBF), pages 1–6...

    2020

  54. [61]

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simon- celli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image process- ing, 13(4):600–612, 2004

    2004

  55. [62]

    G. Wolberg. Image morphing: a survey. The visual com- puter, 14(8-9):360–372, 1998

    1998

  56. [63]

    Z. Xiao, K. Kreis, and A. Vahdat. Tackling the generative learning trilemma with denoising diffusion GANs. arXiv preprint arXiv:2112.07804, 2021

    2021 arXiv

  57. [64]

    Yang and Y

    T. Yang and Y . Ying. Auc maximization in the era of big data and ai: A survey. ACM Computing Surveys, 55(8):1–37, 2022

    2022

  58. [65]

    Zhang, R

    H. Zhang, R. Ramachandra, K. Raja, and C. Busch. Gener- alized single-image-based morphing attack detection using deep representations from vision transformer. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1510–1518, 2024

    2024

  59. [66]

    Zhang, S

    H. Zhang, S. Venkatesh, R. Ramachandra, K. Raja, N. Damer, and C. Busch. MIPGAN—generating strong and high quality morphing attacks using identity prior driven GAN. IEEE Transactions on Biometrics, Behavior, and Iden- tity Science, 3(3):365–383, 2021

    2021

  60. [68]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a percep- tual metric. In Proceedings of the IEEE conference on com- puter vision and pattern recognition, pages 586–595, 2018

    2018

  61. [69]

    Zope and S

    B. Zope and S. B. Zope. A survey of morphing techniques. International Journal of Advanced Engineering, Manage- ment and Science, 3(2):239773, 2017

    2017

Pith tools

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