pith. sign in

arxiv: 2605.25442 · v1 · pith:ROT2BZM4new · submitted 2026-05-25 · 💻 cs.CV

Enhancing Single-Image Facial Demorphing using Multimodal Large Language Models

Nitish Shukla , Arun Ross This is my paper

Pith reviewed 2026-06-29 22:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords facial demorphingmultimodal large language modelsdiffusion modelsface morphing attacksreference-free reconstructionconditional image generation
0
0 comments X

The pith

Semantic embeddings from intermediate MLLM layers condition a coupled RGB diffusion model to enable reference-free facial demorphing.

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

The paper seeks to recover the two original face images from a single morphed composite without any reference photos of the individuals. It proposes doing this by pulling semantic embeddings from the hidden layers of a multimodal large language model and feeding those embeddings directly into a denoising diffusion model that generates both faces jointly in pixel space. The approach avoids the information loss that occurs when first generating text descriptions and then re-encoding them. If correct, the method would let forensic systems reconstruct separate identities from morphing attacks on face recognition. Ablation results are presented to show advantages from middle layers, RGB-domain operation, and full multimodal embeddings over ViT features.

Core claim

By extracting semantic embeddings from intermediate MLLM layers to condition a coupled conditional generation problem, a denoising diffusion model operating directly in the RGB domain can synthesize both constituent faces from a morphed image. This bypasses lossy text generation-reencoding cycles, preserves fine-grained perceptual details such as hair and textures, ensures inter-identity consistency, and requires no reference images or identity overlap between training and test sets.

What carries the argument

Direct use of intermediate MLLM hidden states as conditioning signals for a coupled RGB-domain denoising diffusion model that performs joint synthesis of both identities.

If this is right

  • Middle MLLM layers encode more identity-discriminative representations than early or late layers.
  • RGB-domain demorphing outperforms latent-space approaches by 30-40% at strict operating points.
  • Full MLLM embeddings outperform raw ViT features due to enhanced semantic structuring from multimodal pretraining.
  • The method functions without assuming identity overlap between training and test sets.

Where Pith is reading between the lines

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

  • The direct embedding approach may apply to other image decomposition tasks where semantic guidance is needed without intermediate text.
  • Joint synthesis in pixel space could extend to paired generation problems that require consistency between outputs.
  • If middle-layer embeddings prove generally useful, similar layer-selection strategies might improve conditioning in other diffusion-based restoration methods.

Load-bearing premise

Semantic embeddings extracted from intermediate MLLM layers encode identity-discriminative facial attributes and cues that meaningfully complement low-level pixel information and can condition the diffusion process without reference images or identity overlap between training and test sets.

What would settle it

An experiment that replaces the MLLM hidden-state conditioning with either random vectors or text-generated descriptions and measures whether the 30-40% gain at strict operating points disappears would settle the claim.

Figures

Figures reproduced from arXiv: 2605.25442 by Arun Ross, Nitish Shukla.

Figure 1
Figure 1. Figure 1: Overview of the morphing and demorphing process. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Reconstruction results from our LLM-guided demor [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed prompt-driven demorphing method. A morphed face image is provided to a Multimodal Large Language Model (MLLM) to generate a semantic description, from which hidden states are extracted from intermediate transformer layers. These hidden representations condition a diffusion-based image-to-image model, which takes the morphed image as input and reconstructs the constituent source ima… view at source ↗
Figure 4
Figure 4. Figure 4: Reconstruction results of our method on samples [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

Face recognition systems are increasingly vulnerable to morphing attacks, where a composite image is crafted to match multiple identities, enabling unauthorized access and identity fraud. Existing detection methods identify morphed images but cannot recover constituent images or identities, limiting their forensic utility. This paper presents a novel reference-free facial demorphing framework that leverages Multimodal Large Language Models (MLLMs) to guide a coupled diffusion-based reconstruction process. Our key innovation lies in extracting semantic embeddings from intermediate MLLM layers to condition the demorphing, providing high-level reasoning about facial attributes and identity cues that complement low-level pixel information. We formulate demorphing as a coupled conditional generation problem, where both constituent faces are synthesized jointly through a denoising diffusion model operating directly in the RGB domain, ensuring inter-identity consistency while preserving fine-grained perceptual details. Unlike prior approaches that rely on compressed latent representations or assume identity overlap between training and testing sets, our method bypasses lossy text generation-reencoding cycles by directly utilizing MLLM hidden states as conditioning signals, enabling the denoising network to attend to subtle visual cues such as hair, background, and facial textures. Ablation studies further reveal that middle MLLM layers encode more identity-discriminative representations, RGB-domain demorphing outperforms latent-space approaches by 30--40\% at strict operating points, and full MLLM embeddings provide substantial advantages over raw ViT features through enhanced semantic structuring from multimodal pretraining.

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 present a reference-free facial demorphing method that extracts semantic embeddings from intermediate layers of Multimodal Large Language Models (MLLMs) to directly condition a coupled RGB-domain denoising diffusion model. This enables joint synthesis of the two constituent faces from a single morphed input image, bypassing text generation-reencoding cycles, with ablations indicating that middle MLLM layers are optimal, RGB-domain operation outperforms latent-space baselines by 30-40% at strict operating points, and MLLM embeddings outperform raw ViT features due to multimodal pretraining.

Significance. If the empirical performance claims and the disentanglement capability hold under scrutiny, the work would offer a meaningful advance in forensic tools for morphing attacks by removing the need for reference images or identity overlap between train and test sets. The direct use of MLLM hidden states as conditioning signals, rather than compressed latents or generated text, represents a potentially useful direction for integrating high-level semantic reasoning into pixel-level generative models for biometrics.

major comments (2)
  1. [Abstract] Abstract: The central performance claims (30--40% gains over latent baselines, middle layers best, MLLM > ViT) and ablation outcomes are asserted without any accompanying quantitative tables, figures, dataset details, training protocols, metrics (e.g., verification rates, perceptual distances), error bars, or statistical tests. These omissions make the load-bearing empirical assertions unverifiable from the manuscript.
  2. [Abstract] Abstract and method description: The single shared MLLM embedding extracted from the morphed image is used to condition joint synthesis of two distinct identities, yet no mechanism is described for how this embedding is split, attended, or disentangled across the two output faces in the coupled diffusion process. This is load-bearing for the reference-free claim, as the skeptic concern that the embedding may primarily capture averaged features is not addressed by any ablation or architectural detail.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claims (30--40% gains over latent baselines, middle layers best, MLLM > ViT) and ablation outcomes are asserted without any accompanying quantitative tables, figures, dataset details, training protocols, metrics (e.g., verification rates, perceptual distances), error bars, or statistical tests. These omissions make the load-bearing empirical assertions unverifiable from the manuscript.

    Authors: The full manuscript (Sections 4 and 5) contains the supporting ablation tables, performance curves, dataset descriptions, training protocols, and metrics with error bars. The abstract summarizes these results at a high level. To improve verifiability, we will revise the abstract to incorporate key quantitative highlights (e.g., specific verification rate improvements and layer comparisons) while retaining brevity. revision: yes

  2. Referee: [Abstract] Abstract and method description: The single shared MLLM embedding extracted from the morphed image is used to condition joint synthesis of two distinct identities, yet no mechanism is described for how this embedding is split, attended, or disentangled across the two output faces in the coupled diffusion process. This is load-bearing for the reference-free claim, as the skeptic concern that the embedding may primarily capture averaged features is not addressed by any ablation or architectural detail.

    Authors: The coupled diffusion formulation (Section 3) applies the shared embedding via cross-attention to two parallel denoising branches, with the joint training objective encouraging identity-specific feature extraction. We agree the description of disentanglement is insufficient and the averaging concern is not explicitly ablated. We will add an architectural diagram, expanded conditioning details, and a targeted ablation on embedding splitting in the revised method section. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on ablations, not derivations or self-citations.

full rationale

The paper describes an MLLM-conditioned RGB diffusion model for reference-free demorphing, with all performance claims (30-40% gains, layer selection benefits) presented as outcomes of ablation experiments rather than quantities derived from equations or prior self-citations. No mathematical derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described method. The central conditioning step is justified by empirical results on identity-discriminative embeddings, not by construction or imported uniqueness theorems. This is the common case of a self-contained empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only the abstract is available; the ledger therefore records the modeling assumptions explicitly invoked in the abstract text. No free parameters, axioms, or invented entities are numerically specified.

axioms (2)
  • domain assumption Intermediate MLLM layers encode identity-discriminative representations that complement pixel-level information.
    Invoked in the abstract as the justification for using hidden states rather than final text output or raw ViT features.
  • domain assumption Joint synthesis of both constituent faces through a single denoising process preserves inter-identity consistency.
    Stated as the formulation of demorphing as a coupled conditional generation problem.

pith-pipeline@v0.9.1-grok · 5787 in / 1523 out tokens · 29960 ms · 2026-06-29T22:47:52.701008+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

76 extracted references · 4 canonical work pages · 3 internal anchors

  1. [1]

    Elasticface: Elastic margin loss for deep face recognition,

    F. Boutros, N. Damer, F. Kirchbuchner, and A. Kuijper, “Elasticface: Elastic margin loss for deep face recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  2. [2]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  3. [3]

    50 years of automated face recogni- tion,

    M. Kim, A. Jain, and X. Liu, “50 years of automated face recogni- tion,”IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–20, 2026

    2026

  4. [4]

    The magic passport,

    M. Ferrara, A. Franco, and D. Maltoni, “The magic passport,” inProceedings of International Joint Conference on Biometrics (IJCB), 2014

    2014

  5. [5]

    Unveiling the two-faced truth: Disentangling morphed identities for face morphing detection,

    E. Caldeira, P . C. Neto, T. Gonc ¸alves, N. Damer, A. F. Sequeira, and J. S. Cardoso, “Unveiling the two-faced truth: Disentangling morphed identities for face morphing detection,” inProceedings of European Signal Processing Conference (EUSIPCO), 2023

    2023

  6. [6]

    Mordiff: Recognition vulnerability and attack detectability of face morphing attacks created by diffusion autoencoders,

    N. Damer, M. Fang, P . Siebke, J. N. Kolf, M. Huber, and F. Boutros, “Mordiff: Recognition vulnerability and attack detectability of face morphing attacks created by diffusion autoencoders,” inProceed- ings of International Workshop on Biometrics and Forensics (IWBF), 2023

    2023

  7. [7]

    MIPGAN - generating strong and high quality morphing attacks using identity prior driven GAN,

    H. Zhang, S. Venkatesh, R. Ramachandra, K. B. Raja, N. Damer, and C. Busch, “MIPGAN - generating strong and high quality morphing attacks using identity prior driven GAN,”IEEE Trans- actions on Biometrics, Behavior, and Identity Science (T-BIOM), vol. 3, no. 3, pp. 365–383, 2021

    2021

  8. [8]

    Automatic gener- ation and detection of visually faultless facial morphs,

    A. Makrushin, T. Neubert, and J. Dittmann, “Automatic gener- ation and detection of visually faultless facial morphs,” inPro- ceedings of International Conference on Computer Vision Theory and Applications (VISAPP), 2017

    2017

  9. [9]

    Facemorpher,

    A. Quek, “Facemorpher,” 2019

    2019

  10. [10]

    Training generative adversarial networks with limited data,

    T. Karras, M. Aittala, J. Hellsten, S. Laine, J. Lehtinen, and T. Aila, “Training generative adversarial networks with limited data,” in Proceedings of Conference on Neural Information Processing Systems (NeurIPS), 2020

    2020

  11. [11]

    Ladimo: Face morph generation through biometric template inversion with latent diffusion,

    M. Grimmer and C. Busch, “Ladimo: Face morph generation through biometric template inversion with latent diffusion,” in Proceedings of IEEE International Joint Conference on Biometrics (IJCB), 2024

    2024

  12. [12]

    FLOWING: Implicit neural flows for structure-preserving morphing,

    A. Bizzi, M. G. Portnoy, V . P . Matias, D. Perazzo, J. P . S. do Monte Lima, L. Velho, N. Gonc ¸alves, J. M. Pereira, G. Schardong, and T. Novello, “FLOWING: Implicit neural flows for structure-preserving morphing,” inProceedings of Annual Con- ference on Neural Information Processing Systems (NeurIPS), 2025

    2025

  13. [13]

    Face morphing detection using fourier spectrum of sensor pattern noise,

    L.-B. Zhang, F. Peng, and M. Long, “Face morphing detection using fourier spectrum of sensor pattern noise,” inProceedings of International Conference on Multimedia and Expo (ICME), 2018

    2018

  14. [14]

    Single Im- age Face Morphing Attack Detection Using Ensemble of Features,

    S. Venkatesh, R. Ramachandra, K. Raja, and C. Busch, “Single Im- age Face Morphing Attack Detection Using Ensemble of Features,” inProceedings of 23rd International Conference on Information Fusion (FUSION), 2020

    2020

  15. [15]

    Transfer- able Deep-CNN Features for Detecting Digital and Print-Scanned Morphed Face Images,

    R. Raghavendra, K. B. Raja, S. Venkatesh, and C. Busch, “Transfer- able Deep-CNN Features for Detecting Digital and Print-Scanned Morphed Face Images,” inProceedings of IEEE Conference on Com- puter Vision and Pattern Recognition Workshops (CVPRW), 2017

    2017

  16. [16]

    MADation: Face Morph- ing Attack Detection with Foundation Models,

    E. Caldeira, G. Ozgur, T. Chettaoui, M. Ivanovska, P . Peer, F. Boutros, V . Struc, and N. Damer, “MADation: Face Morph- ing Attack Detection with Foundation Models,” inProceedings of IEEE/CVF Winter Conference on Applications of Computer Vision, WACV Workshops, 2025

    2025

  17. [17]

    Orthomad: Morphing attack detection through or- thogonal identity disentanglement,

    P . C. Neto, T. Gonc ¸alves, M. Huber, N. Damer, A. F. Sequeira, and J. S. Cardoso, “Orthomad: Morphing attack detection through or- thogonal identity disentanglement,” inProceedings of International Conference of the Biometrics Special Interest Group (BIOSIG), 2022

    2022

  18. [18]

    Detecting face morphing attacks by analyzing the directed distances of facial landmarks shifts,

    N. Damer, V . Boller, Y. Wainakh, F. Boutros, P . Terh¨orst, A. Braun, and A. Kuijper, “Detecting face morphing attacks by analyzing the directed distances of facial landmarks shifts,”Pattern Recognition, pp. 518–534, 2019

    2019

  19. [19]

    Detecting morphed face images using facial landmarks,

    U. Scherhag, D. Budhrani, M. Gomez-Barrero, and C. Busch, “Detecting morphed face images using facial landmarks,” inPro- ceedings of International Conference on Image and Signal Processing (ICISP), 2018

    2018

  20. [20]

    Robust morph-detection at automated border control gate using deep decomposed 3d shape & diffuse reflectance,

    J. M. Singh, R. Ramachandra, K. B. Raja, and C. Busch, “Robust morph-detection at automated border control gate using deep decomposed 3d shape & diffuse reflectance,” inProceedings of International Conference on Signal-Image Technology Internet-Based Systems (SITIS), 2019

    2019

  21. [22]

    Fd-gan: Face de-morphing generative adversarial network for restoring accomplice’s facial image,

    F. Peng, L.-B. Zhang, and M. Long, “Fd-gan: Face de-morphing generative adversarial network for restoring accomplice’s facial image,”IEEE Access, vol. 7, pp. 75122–75131, 2019

    2019

  22. [23]

    Face de-morphing based on diffusion autoencoders,

    M. Long, Q. Yao, L.-B. Zhang, and F. Peng, “Face de-morphing based on diffusion autoencoders,”IEEE Transactions on Information Forensics and Security, vol. 19, pp. 3051–3063, 2024

    2024

  23. [24]

    diffDemorph: Extending Reference-Free Demorphing to Unseen Faces,

    N. Shukla and A. Ross, “diffDemorph: Extending Reference-Free Demorphing to Unseen Faces,” inProceedings of IEEE International Conference on Image Processing (ICIP), 2025

    2025

  24. [25]

    Facial demorphing via identity preserving image decomposition,

    N. Shukla and A. Ross, “Facial demorphing via identity preserving image decomposition,” inProceedings of the IEEE International Joint Conference on Biometrics (IJCB), 2024

    2024

  25. [26]

    Facial demorphing from a single morph using a latent conditional gan,

    N. Shukla and A. Ross, “Facial demorphing from a single morph using a latent conditional gan,” inProceedings of IEEE International Joint Conference on Biometrics (IJCB), 2025

    2025

  26. [27]

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

    N. Shukla and A. Ross, “dc-GAN: Dual-Conditioned GAN for Face Demorphing From a Single Morph,” inProceedings of IEEE International Conference on Automatic Face and Gesture Recognition (FG), 2025

    2025

  27. [28]

    Facial De-morphing: Extract- ing Component Faces from a Single Morph,

    S. Banerjee, P . Jaiswal, and A. Ross, “Facial De-morphing: Extract- ing Component Faces from a Single Morph,” inProceedings of IEEE International Joint Conference on Biometrics, 2022

    2022

  28. [29]

    Two algorithms for constructing a delaunay triangulation,

    D. T. Lee and B. J. Schachter, “Two algorithms for constructing a delaunay triangulation,”International Journal of Computer & Information Sciences, vol. 9, pp. 219–242, Jun 1980

    1980

  29. [30]

    GNU Image Manipulation Pro- gram (GIMP), Version 3.0.6. Community, Free Software (license GPLv3),

    GIMP Development Team, “GNU Image Manipulation Pro- gram (GIMP), Version 3.0.6. Community, Free Software (license GPLv3),” 2025. Version 3.0.6, Free Software

    2025

  30. [31]

    GIMP Animation Package,

    The GIMP Development Team, “GIMP Animation Package,” 2025

    2025

  31. [32]

    Active shape models with SIFT descriptors and MARS,

    S. Milborrow and F. Nicolls, “Active shape models with SIFT descriptors and MARS,” inProceedings of International Conference on Computer Vision Theory and Applications, 2014

    2014

  32. [33]

    Face morph using opencv — c++ / python,

    S. Mallick, “Face morph using opencv — c++ / python,” LearnOpenCV, vol. 1, no. 1, 2016

    2016

  33. [34]

    Privacy-friendly Synthetic Data for the Development of Face Morphing Attack Detectors,

    N. Damer, C. A. F. L ´opez, M. Fang, N. Spiller, M. V . Pham, and F. Boutros, “Privacy-friendly Synthetic Data for the Development of Face Morphing Attack Detectors,” inProceedings of Computer Vision and Pattern Recognition Workshops (CVPRW), 2022

    2022

  34. [35]

    Dlib-ml: A Machine Learning Toolkit,

    D. E. King, “Dlib-ml: A Machine Learning Toolkit,”Journal of Machine Learning Research, vol. 10, p. 1755–1758, 2009

    2009

  35. [36]

    Morgan: Recognition vulnerability and attack detectability of face morph- ing attacks created by generative adversarial network,

    N. Damer, A. M. Saladie, A. Braun, and A. Kuijper, “Morgan: Recognition vulnerability and attack detectability of face morph- ing attacks created by generative adversarial network,” inProceed- ings of International Conference on Biometrics Theory, Applications and Systems (BTAS), 2018

    2018

  36. [37]

    Can GAN generated morphs threaten face recog- nition systems equally as landmark based morphs? - vulnerability and detection,

    S. Venkatesh, H. Zhang, R. Ramachandra, K. B. Raja, N. Damer, and C. Busch, “Can GAN generated morphs threaten face recog- nition systems equally as landmark based morphs? - vulnerability and detection,” inProceedings of the International Workshop on Biometrics and Forensics (IWBF), 2020

    2020

  37. [38]

    Pw-mad: Pixel-wise supervision for generalized face morphing attack detection,

    N. Damer, N. Spiller, M. Fang, F. Boutros, F. Kirchbuchner, and A. Kuijper, “Pw-mad: Pixel-wise supervision for generalized face morphing attack detection,”Advances in Visual Computing, pp. 291–304, 2021

    2021

  38. [39]

    Unsupervised face morphing attack detection via self-paced anomaly detection,

    M. Fang, F. Boutros, and N. Damer, “Unsupervised face morphing attack detection via self-paced anomaly detection,” inProceedings of International Joint Conference on Biometrics (IJCB), 2022

    2022

  39. [40]

    SYN-MAD 2022: Competition on face morphing attack detection based on privacy-aware synthetic training data,

    M. Huber, F. Boutros, A. T. Luu, K. Raja, R. Ramachandra, N. Damer, P . C. Neto, T. Gonc ¸alves, A. F. Sequeira, J. S. Cardoso, et al., “SYN-MAD 2022: Competition on face morphing attack detection based on privacy-aware synthetic training data,” inPro- ceedings of IEEE International Joint Conference on Biometrics (IJCB), 2022

    2022

  40. [41]

    Detecting face morphing attacks with collaborative representation of steer- able features,

    R. Ramachandra, S. Venkatesh, K. Raja, and C. Busch, “Detecting face morphing attacks with collaborative representation of steer- able features,” inProceedings of International Conference on Computer Vision and Image Processing (CVIP), 2019. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 14

    2019

  41. [42]

    MSA-CNN: Face Morphing Detection via a Multiple Scales Attention Convolutional Neural Network,

    L.-B. Zhang, J. Cai, F. Peng, and M. Long, “MSA-CNN: Face Morphing Detection via a Multiple Scales Attention Convolutional Neural Network,” inProceedings of 20th International Workshop on Digital Forensics and Watermarking (IWDW), 2021

    2021

  42. [43]

    A Multi-detector Solution Towards an Accurate and Generalized Detection of Face Morphing Attacks,

    N. Damer, S. Zienert, Y. Wainakh, A. M. Saladi ´e, F. Kirchbuchner, and A. Kuijper, “A Multi-detector Solution Towards an Accurate and Generalized Detection of Face Morphing Attacks,” inProceed- ings of 22th International Conference on Information Fusion (FUSION), 2019

    2019

  43. [44]

    Towards making Morphing Attack Detection robust using hybrid Scale- Space Colour Texture Features,

    R. Ramachandra, S. Venkatesh, K. Raja, and C. Busch, “Towards making Morphing Attack Detection robust using hybrid Scale- Space Colour Texture Features,” inProceedings of 5th International Conference on Identity, Security, and Behavior Analysis (ISBA), 2019

    2019

  44. [45]

    General- ized Single-Image-Based Morphing Attack Detection Using Deep Representations from Vision Transformer,

    H. Zhang, R. Ramachandra, K. Raja, and C. Busch, “General- ized Single-Image-Based Morphing Attack Detection Using Deep Representations from Vision Transformer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024

    2024

  45. [46]

    Face morphing attack detection with denoising diffusion probabilistic models,

    M. Ivanovska and V . Struc, “Face morphing attack detection with denoising diffusion probabilistic models,” inProceedings of International Workshop on Biometrics and Forensics (IWBF), 2023

    2023

  46. [47]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,” inProceedings of Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  47. [48]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clark,et al., “Learning transferable visual models from natural language supervision,” in Proceedings of International Conference on Machine Learning (ICML), 2021

    2021

  48. [49]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P . J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,”Journal of Machine Learning Research, vol. 21, Jan. 2020

    2020

  49. [50]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” in Proceedings of NeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021

    2021

  50. [51]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  51. [52]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y. J. Lee, “Visual instruction tuning,” in Proceedings of International Conference on Neural Information Process- ing Systems (NeurIPS), 2023

    2023

  52. [53]

    InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning,

    W. Dai, J. Li, D. Li, A. Tiong, J. Zhao, W. Wang, B. Li, P . Fung, and S. Hoi, “InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning,” inProceedings of Annual Confer- ence on Neural Information Processing Systems (NeurIPS), 2023

    2023

  53. [54]

    GPT-4 technical report,

    OpenAI, “GPT-4 technical report,” 2023

    2023

  54. [55]

    Claude: A new AI assistant,

    Anthropic, “Claude: A new AI assistant,” 2023

    2023

  55. [56]

    Gemini: Advancements in AI models.,

    Google, “Gemini: Advancements in AI models.,” 2023

    2023

  56. [57]

    Sigmoid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sigmoid loss for language image pre-training,” inProceedings of IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2023

    2023

  57. [58]

    Llava-onevision-chat: Improving chat with preference learning,

    T. Xiong, B. Li, D. Guo, H. Yuan, Q. Gu, and C. Li, “Llava-onevision-chat: Improving chat with preference learning,” September 2024

    2024

  58. [59]

    Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization

    W. Wang, Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, J. Zhu, X. Zhu, L. Lu, Y. Qiao,et al., “Enhancing the reasoning ability of multi- modal large language models via mixed preference optimization,” arXiv preprint arXiv:2411.10442, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  59. [60]

    Face demorphing,

    M. Ferrara, A. Franco, and D. Maltoni, “Face demorphing,”IEEE Transactions on Information Forensics and Security, vol. 13, no. 4, pp. 1008–1017, 2017

    2017

  60. [61]

    Border control morphing attack detection with a convolu- tional neural network de-morphing approach,

    D. Ortega-Delcampo, C. Conde, D. Palacios-Alonso, and E. Ca- bello, “Border control morphing attack detection with a convolu- tional neural network de-morphing approach,”IEEE Access, vol. 8, pp. 92301–92313, 2020

    2020

  61. [62]

    MorphGANformer: Transformer-based face morphing and de-morphing,

    N. Zhang, X. Liu, X. Li, and G.-J. Qi, “MorphGANformer: Transformer-based face morphing and de-morphing,”arXiv preprint arXiv:2302.09404, 2023

    work page arXiv 2023

  62. [63]

    A styleGAN- based face de-morphing network for restoring accomplice’s facial image,

    J. Cai, M. Long, L. Zhang, Q. Yao, and X. Ding, “A styleGAN- based face de-morphing network for restoring accomplice’s facial image,”Multimedia Systems, vol. 31, 2025

    2025

  63. [64]

    SDeMorph: Towards Better Facial De-morphing from Single Morph,

    N. Shukla, “SDeMorph: Towards Better Facial De-morphing from Single Morph,” inProceedings of IEEE International Joint Conference on Biometrics (IJCB), 2023

    2023

  64. [65]

    Metric for Evaluating Performance of Reference-Free Demorphing Methods,

    N. Shukla and A. Ross, “Metric for Evaluating Performance of Reference-Free Demorphing Methods,” inProceedings of IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2025

    2025

  65. [66]

    Diffusers: State-of-the-art diffusion models

    P . von Platen, S. Patil, A. Lozhkov, P . Cuenca, N. Lambert, K. Rasul, M. Davaadorj, D. Nair, S. Paul, W. Berman, Y. Xu, S. Liu, and T. Wolf, “Diffusers: State-of-the-art diffusion models.” https://github.com/huggingface/diffusers, 2022

    2022

  66. [67]

    Joint face detection and alignment using multitask cascaded convolutional networks,

    K. Zhang, Z. Zhang, Z. Li, and Y. Qiao, “Joint face detection and alignment using multitask cascaded convolutional networks,” IEEE Signal Processing Letters, vol. 23, pp. 1499–1503, Oct 2016

    2016

  67. [68]

    Extended StirTrace Benchmarking of Biometric and Forensic Qualities of Morphed Face Images,

    T. Neubert, A. Makrushin, M. Hildebrandt, C. Kraetzer, and J. Dittmann, “Extended StirTrace Benchmarking of Biometric and Forensic Qualities of Morphed Face Images,”IET Biometrics, vol. 7, pp. 325–332, 2018

    2018

  68. [69]

    Face Research Lab London Set,

    L. DeBruine and B. Jones, “Face Research Lab London Set,” 5 2017

    2017

  69. [70]

    debruine/webmorph: Beta release 2,

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

    2018

  70. [71]

    Qwen3 Technical Report

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M. Li, M. Xue, M. Li, P . Zhang, P . Wang, Q. Zhu, R. Men, R. Gao, S. Liu, S. Luo, T...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  71. [72]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y. Li, and Y. J. Lee, “Improved baselines with visual instruction tuning,” 2023

    2023

  72. [73]

    Adam: A Method for Stochastic Op- timization,

    D. P . Kingma and J. Ba, “Adam: A Method for Stochastic Op- timization,” inProceedings of International Conference on Learning Representations (ICLR), 2017

    2017

  73. [74]

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

    S. Gugger, L. Debut, T. Wolf, P . Schmid, Z. Mueller, S. Mangrulkar, M. Sun, and B. Bossan, “Accelerate: Training and inference at scale made simple, efficient and adaptable.” https://github.com/ huggingface/accelerate, 2022

    2022

  74. [75]

    Adaface: Quality adaptive margin for face recognition,

    M. Kim, A. K. Jain, and X. Liu, “Adaface: Quality adaptive margin for face recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  75. [76]

    Transformers: State-of- the-art natural language processing,

    T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P . Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P . von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. L. Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush, “Transformers: State-of- the-art natural language processing,” inProceedings of the Confer- ence on Empirical Meth...

    2020

  76. [77]

    Learning Face Representation from Scratch

    D. Yi, Z. Lei, S. Liao, and S. Z. Li, “Learning face representation from scratch,”arXiv preprint arXiv:1411.7923, 2014. Nitish Shuklais currently a Ph.D. candidate in the Department of Computer Science and Engineering at Michigan State University. His research interests include facial biometrics and multimodal learning. Arun Ross(Senior Member, IEEE) is t...

    work page internal anchor Pith review Pith/arXiv arXiv 2014