Pith. sign in

REVIEW 4 major objections 4 minor 91 references

Vec2Face+ for Face Dataset Generation

T0 review · 4 major / 4 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 Synthetic face images now beat a real training set.

desk verdict A well-engineered synthetic face dataset paper whose headline milestone is fragile: 0.1 pp margins, single runs, and a scale mismatch, but the twin/bias diagnostics and code are worth engaging. read the letter →

arxiv 2507.17192 v2 pith:VSEVWG4H submitted 2025-07-23 cs.CV

Haiyu Wu , Jaskirat Singh , Sicong Tian , Liang Zheng , Kevin W. Bowyer This is my paper · ORCID
classification cs.CV
keywords syntheticfacedatasetgenerationrecognitionintra-classidentityconsistencyidentity-conditionedfeaturemaskedautoencoderArcembeddingsprivacy-preservingtrainingdatatwinsverification
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 claims that the missing ingredient in synthetic face training data is intra-class identity consistency: existing generators produce images with plenty of attribute variation and well-separated identities, but images meant to show one person often drift into different people, injecting noise into training. To fix this, the authors build Vec2Face+, a generator that creates images directly from face-recognition embeddings and adds LoRA-based pose control and a gradient-based attribute operator. With it they generate VFace datasets at 10K, 20K, 100K, and 300K identities. The headline result is that VFace100K and VFace300K reach 94.88% and 94.93% average accuracy on five real-world test sets, slightly above the 94.79% of CASIA-WebFace, the first time a synthetic dataset has outperformed that real training set in average accuracy. The paper also reports two negative findings: nearly all synthetic datasets fail to beat random guessing on identical-twin verification, and models trained on synthetic identities show larger demographic disparities.

What carries the argument

The engine is a feature masked auto-encoder (fMAE) generator that maps 512-dimensional face-recognition embeddings directly to images: the embedding is expanded to a feature map, rows are randomly dropped before encoding, the dropped positions are filled with a conditioning signal (the image feature, or a CNN-extracted face-landmark feature for pose), and a four-layer deconvolutional decoder renders the 112x112 image. Identity is carried by the embedding itself, so inter-class separability is controlled by sampling vectors with cosine similarity at most 0.3, and intra-class variation comes from Gaussian perturbations of the identity vector, from the AttrOP gradient-descent algorithm that pushes generated images toward target pose and quality values, and from LoRA fine-tuning conditioned on five face landmark points for profile poses. The identity loss, reconstruction loss, and perceptual loss jointly supervise reconstruction in a way that keeps generated images close to their source embedding in ArcFace space, which is also the space where dataset consistency is measured.

What would settle it

Train the same downstream matcher on the VFace datasets but measure accuracy and intra-class consistency using a different identity representation, for example a ViT-based embedding or human similarity ratings; if VFace's edge over CASIA-WebFace vanishes or consistency no longer tracks accuracy, the central claim is tied to the ArcFace space. Independently, running the VFace300K-versus-CASIA-WebFace comparison across several training seeds would show whether the 0.14-point advantage is within run-to-run variance.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that intra-class identity consistency, not intra-class attribute variation or inter-class separability, is the factor holding synthetic face training sets below real ones. It diagnoses this by measuring nine attribute distributions, inter-class separability, and the mean pairwise cosine similarity of images within identity folders across six datasets, finding that synthetic sets match or exceed real data on attribute variation and separability but fall short on within-identity consistency. It then proposes Vec2Face+, which generates images directly from ArcFace identity embeddings and controls attributes through Gaussian perturbation, a gradient-descent attribute operator, and landmark-conditioned LoRA pose control. The resulting VFace100K and VFace300K datasets are the first synthetic training sets reported to beat the real-world CASIA-WebFace in average accuracy on five standard face benchmarks, while VFace10K outperforms other same-scale synthetic datasets by 1.59 points on average. The paper also argues that the similarity-based identity definition used throughout synthetic dataset generation causes matchers trained on such data to fail on identical-twin verification and to show larger demographic accuracy gaps.

Load-bearing premise

The load-bearing premise is that the fixed face-recognition embedding used to sample identities, supervise generation, measure dataset consistency, and train the final matcher is a faithful proxy for human identity; if that representation is not general, both the identity-consistency story and the accuracy gains could be partly an artifact of optimizing and evaluating in the same space.

Editorial extensions

If this is right

  • A face matcher trained on synthetic-only data can match or beat one trained on the widely used real dataset CASIA-WebFace on five standard benchmarks, so privacy-preserving synthetic training data need not be a performance sacrifice.
  • Intra-class identity consistency, not attribute diversity or inter-class separability alone, is the bottleneck: datasets with higher within-identity consistency produce higher accuracy, and the benefit of separability saturates around 0.7.
  • The similarity-based identity definition used in current synthetic generation fails on identical-twins verification: 10 of 11 synthetic datasets produce matchers at or below random chance on Twins-IND.
  • Models trained on synthetic identities show larger demographic accuracy gaps than models trained on real data on bias-aware test sets.
  • Scaling the number of synthetic identities from 10K to 300K yields monotonic accuracy gains on standard test sets, showing the generation pipeline supports dataset scaling to 12M images.

Reading between the lines

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

  • A testable extension follows directly: enforce identity consistency in a representation-independent way, for instance using multiple independently trained embeddings or human similarity judgments during generation, and check whether the accuracy edge over CASIA-WebFace survives; if it does not, the causal story is partly an artifact of the ArcFace space.
  • The reported win over CASIA-WebFace is about 0.09 to 0.14 average points, so the headline claim is best read as 'at parity with a real dataset' until multiple independent training runs confirm the gap is stable; the paper does not report seed variance for this comparison.
  • The twin-verification and demographic-bias failures point to a concrete next step: generate identities with explicit kinship structure, such as shared latent factors between some identities, and explicitly balance demographic attributes, then measure whether those two weaknesses close.
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 / 4 minor

Summary. The paper introduces Vec2Face+, a generative model that maps face-recognition feature vectors to face images, together with three dataset-assembly strategies: well-separated identity sampling, AttrOP for attribute variation, and LoRA-based pose control. The authors generate VFace datasets with 10K, 20K, 100K, and 300K identities, report that VFace10K is the best-performing synthetic training set among the compared methods, and claim that VFace100K and VFace300K are the first synthetic datasets to exceed CASIA-WebFace in average accuracy on five standard face-verification test sets. The paper also analyzes attribute variation, inter-class separability, intra-class identity consistency, twin-verification failure, and demographic bias of synthetic training data.

Significance. If the headline result holds, this is an important step toward privacy-preserving face-recognition training, and the paper's extensive analysis of attribute variation, identity consistency, and failure modes (twins, demographic bias) is a genuine contribution. The release of code, the explicit admission that synthetic datasets fail at twin verification, and the human-evaluation experiment all increase the transparency of the work. However, the central 'first synthetic dataset beats CASIA-WebFace' claim rests on a per-dataset average margin of 0.09–0.14 percentage points from single training runs, and the entire pipeline and evaluation live in the same ArcFace-R100 feature space; both issues must be resolved before the claim can be accepted.

major comments (4)
  1. [V-A / Table I] The headline claim that VFace100K and VFace300K outperform CASIA-WebFace rests on average-accuracy margins of +0.09 and +0.14 percentage points in Table I. Section V-A reports a single training run per dataset with no random seeds or confidence intervals; for 40-epoch SGD face-recognition training, this margin is within the range of seed-to-seed variation. Consequently the statement 'This is the first time a synthetic dataset beats CASIA-WebFace in average accuracy' is not yet supported. Please provide multi-seed results (mean ± std, with the number of runs) for at least VFace100K, VFace300K, and CASIA-WebFace, and ideally for all entries in Tables I, II, and IV.
  2. [Table I note / V-B] The caption of Table I states that Hadrian, Eclipse, IJBB, and IJBC entries marked with ⋆ come from models trained on the distributed dataset, while Section V-B states that these numbers were reproduced by retraining with the available datasets. This inconsistency means the nine-test-set SOTA claim cannot be independently verified. The Hadrian/Eclipse/IJBB/IJBC comparisons should be rerun under the paper's own protocol or clearly labeled as external and excluded from the comparative claims.
  3. [IV-D] The identity-leakage filter (dropping generated images whose ArcFace similarity to WebFace4M identities exceeds 0.4) is applied only to VFace data, not to any of the baseline synthetic datasets in Table I. This could give VFace an unfair advantage in the comparison. Please state whether the baselines were filtered in the same way; if not, discuss the effect of the filtering on the accuracy comparison.
  4. [III-C / IV-A / V-A] The generation pipeline, the dataset-quality metrics, and the downstream matcher all use the ArcFace-R100 feature space. Identity vectors are sampled in this space, the reconstruction is supervised with ArcFace cosine distance, intra-class consistency is measured by ArcFace cosine similarity, and the trained model uses ArcFace loss. The paper's claim that intra-class identity consistency is 'the missing factor' is thus, in part, a claim about consistency in the ArcFace space rather than about human identity. The paper's own human evaluation (Fig. 15) indicates that VFace is not more identity-consistent than CASIA-WebFace by human judgment, which weakens the causal story. Please add an explicit discussion of this possible circularity and provide a human-aligned identity-consistency evaluation for the proposed method.
minor comments (4)
  1. [Section I] There is a typo in the phrase 'may turn out to be of diffferent identities'; 'diffferent' should be 'different'.
  2. [IV-A] The description of the masked-row distribution as 'Ntruncated(max = 1, min= 0.5, mean= 0.75)' is unclear; please define the exact distribution and its parameters.
  3. [II] Figure 2's characterization of all GAN- and diffusion-based methods as combining 'a Gaussian noise image and conditions' is an oversimplification that may misrepresent recent conditional generative models; consider softening or rewording.
  4. [V-A] Please state explicitly whether all training runs (including the ablations in Table IV) use the same random seed; otherwise the differences in the ablation table could be due to training noise rather than the ablated components.

Circularity Check

1 steps flagged · score 5.0 of 10

Dataset-quality metric is the training objective in the same ArcFace space; the headline benchmark comparison is external and not circular.

  1. self definitional [Section III-C Eq. (2); Section IV-A Eq. (4); Section IV-C Eqs. (8)-(9)]
    "The equation for intra-class identity consistency measurement is: Dconsis = 1/NK Σ_i Σ_j (fi·fj)/(||fi|| ||fj||) ... where N is the number of identities, K is the number of images for an identity, and f is the image feature vector extracted by a pre-trained FR model. ... the identity loss is measured by cosine distance: Lid = 1− frec·fgt/(||frec|| ||fgt||), where the f is the feature vector extracted by a pre-trained FR model."

    The dataset-quality metric Dconsis and the generator's identity loss Lid are the same cosine similarity in the same pre-trained ArcFace-R100 feature space, up to sign and averaging. The pipeline explicitly optimizes Lid (Eq. 4) and filters generated images by Sim(vid,fI)>0.9 and Sim(fI,fP)>0.7 (Eqs. 8-9). VFace's high intra-class consistency reported in Fig. 6 is therefore enforced by the training objective and selection thresholds, not discovered by measurement. The paper then offers this by-construction gain as evidence that 'intra-class identity consistency is the missing puzzle.' The headline accuracy gain over CASIA-WebFace is measured on external test sets and does not reduce to this construction, so the circularity is partial rather than total.

full rationale

The central benchmark result—VFace100K and VFace300K outperform CASIA-WebFace in average accuracy on five standard test sets—is externally evaluated and does not reduce to the paper's own definitions or fitted parameters. It is a legitimate empirical comparison, despite being a single-run result with small margins. The main circularity concern is the supporting causal claim that intra-class identity consistency is the neglected factor. Consistency is defined in Eq. (2) as average cosine similarity in the ArcFace-R100 feature space, and the Vec2Face+ generator is trained with Eq. (4), which is exactly the complement of that cosine, and with constraints that enforce the same similarity in the same space. Thus VFace's consistency advantage over other synthetic sets is a consequence of the optimization objective, not an independent finding. The paper's own human evaluation in Fig. 15 further weakens the external grounding: by human judgment, CASIA-WebFace has better intra-class consistency than VFace, suggesting the machine-space advantage is tied to the feature space in which the data was generated and evaluated. No load-bearing self-citation was found: the Vec2Face basis is transparently stated, and the VFace results are not derived from that citation alone. The accuracy comparison and the twin-verification and demographic-bias findings remain independent empirical contributions, so the circularity score is moderate rather than severe.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

No new physical or theoretical entities are introduced; synthetic identities are generated data rather than new objects. The ledger shows the method depends on hand-set sampling and quality thresholds, and on the ArcFace feature space as the operative definition of identity.

free parameters (8)
  • perceptual loss weight lambda = 0.2
    Hand-set default weighting of LPIPS loss in Eq. (6).
  • perturbation noise sigmas = {0.3, 0.5, 0.7}
    Chosen to create intra-class variation while keeping cosine similarity to identity at least 0.5 (Sec. IV-C).
  • identity vector norm range = 18 to 24
    Normalization range to avoid corrupted face images (Sec. IV-C).
  • identity sampling similarity bound = cosine similarity <= 0.3
    Controls inter-class separability in vector space (Sec. I and IV-C).
  • quality thresholds for generated images = Q>26 identity images, Q>24 perturbed images, Sim>0.9 and Sim>0.7
    Filters used to keep generated identity images and intra-class variants close to target vectors (Eqs. 8 and 9).
  • AttrOP target yaw angles and max iterations = yaw in {30,40,50,60,70,80}, up to 30 iterations
    Chosen to balance pose variation against generation time (Sec. IV-D).
  • feature-map row mask ratio = x in [0.5, 1], mean 0.75
    fMAE masking range fixed by the authors for training (Sec. IV-A).
  • separability threshold = 0.4
    Adopted from Vec2Face [7] to count well-separated identities.
assumptions (4)
  • domain assumption ArcFace-R100 features are a faithful identity representation for generation, dataset analysis, and matcher training.
    Used in Sec. IV-A for supervision and in Sec. III for consistency and separability metrics; if false, the causal attribution is representation-specific.
  • domain assumption The attribute estimators (BiSeNet, MagFace, img2pose, DDAMFN, LogicNet, FairFace) give accurate enough labels for dataset comparisons.
    The analysis in Sec. III-A relies on these pretrained models to measure attribute variation.
  • domain assumption Baseline numbers taken from distributed datasets or prior papers are directly comparable to the uniformly retrained VFace models.
    Table I marks some Hadrian, Eclipse, IJBB, and IJBC entries with a star; those models were not retrained under the paper's exact protocol.
  • domain assumption The identity leakage filter and DBSCAN pruning do not give VFace an unfair advantage over unfiltered baselines.
    Applied only to VFace datasets in Sec. IV-D; baseline datasets are not similarly cleaned.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vec2Face+ for Face Dataset Generation." pith.science (2026). https://pith.science/paper/VSEVWG4H

@misc{pith2026250717192,
  author       = {Pith},
  title        = {Pith review of: Vec2Face+ for Face Dataset Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VSEVWG4H}},
  note         = {Machine review of arXiv:2507.17192}
}
read the original abstract

When synthesizing identities as face recognition training data, it is generally believed that large inter-class separability and intra-class attribute variation are essential for synthesizing a quality dataset. % This belief is generally correct, and this is what we aim for. However, when increasing intra-class variation, existing methods overlook the necessity of maintaining intra-class identity consistency. % To address this and generate high-quality face training data, we propose Vec2Face+, a generative model that creates images directly from image features and allows for continuous and easy control of face identities and attributes. Using Vec2Face+, we obtain datasets with proper inter-class separability and intra-class variation and identity consistency using three strategies: 1) we sample vectors sufficiently different from others to generate well-separated identities; 2) we propose an AttrOP algorithm for increasing general attribute variations; 3) we propose LoRA-based pose control for generating images with profile head poses, which is more efficient and identity-preserving than AttrOP. % Our system generates VFace10K, a synthetic face dataset with 10K identities, which allows an FR model to achieve state-of-the-art accuracy on seven real-world test sets. Scaling the size to 4M and 12M images, the corresponding VFace100K and VFace300K datasets yield higher accuracy than the real-world training dataset, CASIA-WebFace, on five real-world test sets. This is the first time a synthetic dataset beats the CASIA-WebFace in average accuracy. In addition, we find that only 1 out of 11 synthetic datasets outperforms random guessing (\emph{i.e., 50\%}) in twin verification and that models trained with synthetic identities are more biased than those trained with real identities. Both are important aspects for future investigation. Code is available at https://github.com/HaiyuWu/Vec2Face_plus

Figures

Figures reproduced from arXiv: 2507.17192 by the authors.

Figure 1
Figure 1. Comparing existing synthetic FR training sets of their average accuracy [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. GAN-based and diffusion-based methods combine a Gaussian noise [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Does intra-class variation decide training data effectiveness? We compare five synthetic datasets and a real dataset (CAS.), where each figure shows a [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Comparing synthetic and real FR training sets of their inter [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Examples of inconsistent identities in five synthetic datasets. Images [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 7
Figure 7. Figure 7: Architecture of Vec2Face+. The input is the image feature, [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 9
Figure 9. Figure 9: Architecture of conditional fine-tuning. A CNN block and parameter [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 11
Figure 11. Figure 11: Examples of generated images for one identity by three different [PITH_FULL_IMAGE:figures/full_fig_p007_11.png]
Figure 12
Figure 12. Figure 12: Correlation between inter-class separability and test accuracy. From left to right: test on (a) five test sets, including LFW, CFP-FP, CPLFW, AgeDB-30, [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: Correlation between intra-class identity consistency and test accuracy. From left to right: test on (a) five test sets, including LFW, CFP-FP, CPLFW, [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 14
Figure 14. Figure 14: Inter-class separability measured by human vs. test accuracy. We ask human annotators to work on the most similar inter-class pairs in each dataset [PITH_FULL_IMAGE:figures/full_fig_p012_14.png]
Figure 15
Figure 15. Figure 15: Intra-class ID consistency measured by human vs. test accuracy. We ask human annotators to work on the most dissimilar intra-class pairs in each [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

91 extracted references · 65 canonical work pages

  1. [1]

    FRCSyn challenge at CVPR 2024: Face recognition challenge in the era of synthetic data,

    I. DeAndres-Tame, R. Tolosana, P. Melzi, R. Vera-Rodriguez, M. Kim, C. Rathgeb, X. Liu, A. Morales, J. Fierrez, J. Ortega-Garcia et al. , “FRCSyn challenge at CVPR 2024: Face recognition challenge in the era of synthetic data,” in CVPRW, 2024, pp. 3173–3183

    2024

  2. [2]

    FRCSyn challenge at W ACV 2024: Face recognition challenge in the era of synthetic data,

    P. Melzi, R. Tolosana, R. Vera-Rodriguez, M. Kim, C. Rathgeb, X. Liu, I. DeAndres-Tame, A. Morales, J. Fierrez, J. Ortega-Garcia et al. , “FRCSyn challenge at W ACV 2024: Face recognition challenge in the era of synthetic data,” in WACVW, 2024, pp. 892–901

    2024

  3. [3]

    Sdfr: Synthetic data for face recognition competition,

    H. O. Shahreza, C. Ecabert, A. George, A. Unnervik, S. Marcel, N. Di Domenico, G. Borghi, D. Maltoni, F. Boutros, J. V ogel et al. , “Sdfr: Synthetic data for face recognition competition,” in IEEE F&G, 2024, pp. 1–9

    2024

  4. [4]

    50 years of automated face recognition,

    M. Kim, A. Jain, and X. Liu, “50 years of automated face recognition,” arXiv preprint arXiv:2505.24247 , 2025

    work page arXiv 2025

  5. [5]

    DCFace: Synthetic face generation with dual condition diffusion model,

    M. Kim, F. Liu, A. K. Jain, and X. Liu, “DCFace: Synthetic face generation with dual condition diffusion model,” in CVPR. IEEE, 2023, pp. 12 715–12 725

    2023

  6. [6]

    IDiff-face: Synthetic-based face recognition through fizzy identity-conditioned dif- fusion model,

    F. Boutros, J. H. Grebe, A. Kuijper, and N. Damer, “IDiff-face: Synthetic-based face recognition through fizzy identity-conditioned dif- fusion model,” in ICCV, 2023, pp. 19 650–19 661

    2023

  7. [7]

    Vec2Face: Scaling face dataset generation with loosely constrained vectors,

    H. Wu, J. Singh, S. Tian, L. Zheng, and K. W. Bowyer, “Vec2Face: Scaling face dataset generation with loosely constrained vectors,” ICLR, 2025. 13

    2025

  8. [8]

    SynFace: Face recognition with synthetic data,

    H. Qiu, B. Yu, D. Gong, Z. Li, W. Liu, and D. Tao, “SynFace: Face recognition with synthetic data,” in ICCV, 2021, pp. 10 860–10 870

    2021

Show all 91 references
  1. [9]

    SFace: Privacy-friendly and accurate face recognition using synthetic data,

    F. Boutros, M. Huber, P. Siebke, T. Rieber, and N. Damer, “SFace: Privacy-friendly and accurate face recognition using synthetic data,” in IJCB. IEEE, 2022, pp. 1–11

    2022

  2. [10]

    The chroma-fit dataset: Characterizing human ranges of melanin for increased tone-awareness,

    G. Pangelinan, X. Merino, S. Langborgh, K. Vangara, J. Annan, A. Beaubrun, T. Weekes, and M. C. King, “The chroma-fit dataset: Characterizing human ranges of melanin for increased tone-awareness,” in WACVW, 2024, pp. 1170–1178

    2024

  3. [11]

    Face recognition accuracy across demographics: Shining a light into the problem,

    H. Wu, V . Albiero, K. Krishnapriya, M. C. King, and K. W. Bowyer, “Face recognition accuracy across demographics: Shining a light into the problem,” in CVPRW, 2023, pp. 1041–1050

    2023

  4. [12]

    A comparative study on illumination preprocessing in face recognition,

    H. Han, S. Shan, X. Chen, and W. Gao, “A comparative study on illumination preprocessing in face recognition,” PR, vol. 46, no. 6, pp. 1691–1699, 2013

    2013

  5. [13]

    Multi-pie,

    R. Gross, I. Matthews, J. Cohn, T. Kanade, and S. Baker, “Multi-pie,” IVC, vol. 28, no. 5, pp. 807–813, 2010

    2010

  6. [14]

    The cmu pose, illumination, and ex- pression database,

    T. Sim, S. Baker, and M. Bsat, “The cmu pose, illumination, and ex- pression database,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 12, pp. 1615–1618, 2003

    2003

  7. [15]

    Low-resolution face recognition,

    Z. Cheng, X. Zhu, and S. Gong, “Low-resolution face recognition,” in ACCV, 2019, pp. 605–621

    2019

  8. [16]

    Cross-quality LFW: A database for analyzing cross-resolution image face recognition in unconstrained environments,

    M. Knoche, S. Hormann, and G. Rigoll, “Cross-quality LFW: A database for analyzing cross-resolution image face recognition in unconstrained environments,” in IEEE F&G, 2021, pp. 1–5

    2021

  9. [17]

    Logical consistency and greater descriptive power for facial hair attribute learning,

    H. Wu, G. Bezold, A. Bhatta, and K. W. Bowyer, “Logical consistency and greater descriptive power for facial hair attribute learning,” inCVPR, 2023, pp. 8588–8597

    2023

  10. [18]

    Beard segmentation and recognition bias,

    K. Ozturk, G. Bezold, A. Bhatta, H. Wu, and K. Bowyer, “Beard segmentation and recognition bias,” arXiv preprint arXiv:2308.15740 , 2023

    2023 arXiv

  11. [19]

    Can the accuracy bias by facial hairstyle be reduced through balancing the training data?

    K. Ozturk, H. Wu, and K. W. Bowyer, “Can the accuracy bias by facial hairstyle be reduced through balancing the training data?” in CVPRW, 2024, pp. 1519–1528

    2024

  12. [20]

    A comprehensive study on face recognition biases beyond demographics,

    P. Terh ¨orst, J. N. Kolf, M. Huber, F. Kirchbuchner, N. Damer, A. M. Moreno, J. Fierrez, and A. Kuijper, “A comprehensive study on face recognition biases beyond demographics,” IEEE Transactions on Tech- nology and Society , vol. 3, no. 1, pp. 16–30, 2021

    2021

  13. [21]

    The gender gap in face recognition accuracy is a hairy problem,

    A. Bhatta, V . Albiero, K. W. Bowyer, and M. C. King, “The gender gap in face recognition accuracy is a hairy problem,” in WACVW, 2023, pp. 303–312

    2023

  14. [22]

    The AR face database,

    A. Martinez and R. Benavente, “The AR face database,” CVC TechRep #24, 1998

    1998

  15. [23]

    IARPA janus benchmark - C: face dataset and protocol,

    B. Maze, J. C. Adams, J. A. Duncan, N. D. Kalka, T. Miller, C. Otto, A. K. Jain, W. T. Niggel, J. Anderson, J. Cheney, and P. Grother, “IARPA janus benchmark - C: face dataset and protocol,” in ICB, 2018, pp. 158– 165

    2018

  16. [24]

    Impact of sunglasses on one-to-many facial identification accuracy,

    S. Tian, H. Wu, M. C. King, and K. W. Bowyer, “Impact of sunglasses on one-to-many facial identification accuracy,” IEEE FG, 2024

    2024

  17. [25]

    Facial hair area in face recognition across demographics: Small size, big effect,

    H. Wu, S. Tian, A. Bhatta, K. ¨Ozt¨urk, K. Ricanek, and K. W. Bowyer, “Facial hair area in face recognition across demographics: Small size, big effect,” in WACVW, 2024, pp. 1131–1140

    2024

  18. [26]

    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

    2014 arXiv

  19. [27]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, 2022

    2022

  20. [28]

    What is a goldilocks face verification test set?

    H. Wu, S. Tian, A. Bhatta, J. Gutierrez, G. Bezold, G. Argueta, K. Ricanek Jr., M. C. King, and K. W. Bowyer, “What is a goldilocks face verification test set?” arXiv preprint arXiv:2405.15965 , 2024

    2024

  21. [29]

    IARPA Janus benchmark-b face dataset,

    C. Whitelam, E. Taborsky, A. Blanton, B. Maze, J. C. Adams, T. Miller, N. D. Kalka, A. K. Jain, J. A. Duncan, K. Allen, J. Cheney, and P. Grother, “IARPA Janus benchmark-b face dataset,” in CVPRW, 2017, pp. 592–600

    2017

  22. [30]

    Doppelver: A benchmark for face verification,

    N. Thom, A. DeBolt, L. Brown, and E. M. Hand, “Doppelver: A benchmark for face verification,” in ISVC, 2023, pp. 431–444

    2023

  23. [31]

    Face recognition: too bias, or not too bias?

    J. P. Robinson, G. Livitz, Y . Henon, C. Qin, Y . Fu, and S. Timoner, “Face recognition: too bias, or not too bias?” in CVPRW, 2020, pp. 0–1

    2020

  24. [32]

    What should be balanced in a

    H. Wu and K. W. Bowyer, “What should be balanced in a” balanced” face recognition dataset,” in BMVC, 2023, p. 2

    2023

  25. [33]

    Webface260m: A benchmark for million- scale deep face recognition,

    Z. Zhu, G. Huang, J. Deng, Y . Ye, J. Huang, X. Chen, J. Zhu, T. Yang, D. Du, J. Lu, and J. Zhou, “Webface260m: A benchmark for million- scale deep face recognition,” TPAMI, pp. 2627–2644, 2023

    2023

  26. [34]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” in CVPR, 2021, pp. 12 873–12 883

    2021

  27. [35]

    Mage: Masked generative encoder to unify representation learning and image synthesis,

    T. Li, H. Chang, S. Mishra, H. Zhang, D. Katabi, and D. Krishnan, “Mage: Masked generative encoder to unify representation learning and image synthesis,” in CVPR, 2023, pp. 2142–2152

    2023

  28. [36]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in ICCV, 2023, pp. 4195–4205

    2023

  29. [37]

    All are worth words: a VIT backbone for score-based diffusion models,

    F. Bao, C. Li, Y . Cao, and J. Zhu, “All are worth words: a VIT backbone for score-based diffusion models,” in NeurIPS, 2022

    2022

  30. [38]

    Masked diffusion trans- former is a strong image synthesizer,

    S. Gao, P. Zhou, M.-M. Cheng, and S. Yan, “Masked diffusion trans- former is a strong image synthesizer,” inICCV, 2023, pp. 23 164–23 173

    2023

  31. [39]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregressive modeling: Scalable image generation via next-scale prediction,” CVPR, 2024

    2024

  32. [40]

    Fastcomposer: Tuning-free multi-subject image generation with localized attention,

    G. Xiao, T. Yin, W. T. Freeman, F. Durand, and S. Han, “Fastcomposer: Tuning-free multi-subject image generation with localized attention,” arXiv preprint arXiv:2305.10431 , 2023

    2023 arXiv

  33. [41]

    Photoverse: Tuning-free image customization with text-to-image diffusion models,

    L. Chen, M. Zhao, Y . Liu, M. Ding, Y . Song, S. Wang, X. Wang, H. Yang, J. Liu, K. Du et al. , “Photoverse: Tuning-free image customization with text-to-image diffusion models,” arXiv preprint arXiv:2309.05793, 2023

    2023 arXiv

  34. [42]

    Photomaker: Customizing realistic human photos via stacked id embedding,

    Z. Li, M. Cao, X. Wang, Z. Qi, M.-M. Cheng, and Y . Shan, “Photomaker: Customizing realistic human photos via stacked id embedding,” in CVPR, 2024, pp. 8640–8650

    2024

  35. [43]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in ICML, 2021, pp. 8748–8763

    2021

  36. [44]

    Arc2face: A foundation model of human faces,

    F. P. Papantoniou, A. Lattas, S. Moschoglou, J. Deng, B. Kainz, and S. Zafeiriou, “Arc2face: A foundation model of human faces,” arXiv preprint arXiv:2403.11641, 2024

    2024 arXiv

  37. [45]

    Instantid: Zero-shot identity-preserving generation in seconds,

    Q. Wang, X. Bai, H. Wang, Z. Qin, and A. Chen, “Instantid: Zero-shot identity-preserving generation in seconds,” arXiv preprint arXiv:2401.07519, 2024

    2024 arXiv

  38. [46]

    Face0: Instanta- neously conditioning a text-to-image model on a face,

    D. Valevski, D. Lumen, Y . Matias, and Y . Leviathan, “Face0: Instanta- neously conditioning a text-to-image model on a face,” in SIGGRAPH, 2023, pp. 1–10

    2023

  39. [47]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models,

    H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang, “Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models,” arXiv preprint arXiv:2308.06721 , 2023

    2023 arXiv

  40. [48]

    Fake it till you make it: face analysis in the wild using synthetic data alone,

    E. Wood, T. Baltrusaitis, C. Hewitt, S. Dziadzio, T. J. Cashman, and J. Shotton, “Fake it till you make it: face analysis in the wild using synthetic data alone,” in ICCV. IEEE, 2021, pp. 3661–3671

    2021

  41. [49]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in ICCV, 2023, pp. 3836–3847

    2023

  42. [50]

    Unsuper- vised face recognition using unlabeled synthetic data,

    F. Boutros, M. Klemt, M. Fang, A. Kuijper, and N. Damer, “Unsuper- vised face recognition using unlabeled synthetic data,” IEEE F&G, pp. 1–8, 2022

    2022

  43. [51]

    SFace2: Synthetic-based face recognition with w-space identity-driven sam- pling,

    F. Boutros, M. Huber, A. T. Luu, P. Siebke, and N. Damer, “SFace2: Synthetic-based face recognition with w-space identity-driven sam- pling,” IEEE T-BIOM, 2024

    2024

  44. [52]

    ExFace- GAN: Exploring identity directions in GAN’s learned latent space for synthetic identity generation,

    F. Boutros, M. Klemt, M. Fang, A. Kuijper, and N. Damer, “ExFace- GAN: Exploring identity directions in GAN’s learned latent space for synthetic identity generation,” in IJCB. IEEE, 2023, pp. 1–10

    2023

  45. [53]

    Ilvr: Conditioning method for denoising diffusion probabilistic models,

    J. Choi, S. Kim, Y . Jeong, Y . Gwon, and S. Yoon, “Ilvr: Conditioning method for denoising diffusion probabilistic models,” ICCV, 2021

    2021

  46. [54]

    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 CVPR, 2022, pp. 10 684–10 695

    2022

  47. [55]

    DigiFace-1M: 1 million digital face images for face recognition,

    G. Bae, M. de La Gorce, T. Baltrusaitis, C. Hewitt, D. Chen, J. P. C. Valentin, R. Cipolla, and J. Shen, “DigiFace-1M: 1 million digital face images for face recognition,” in WACV, 2023, pp. 3515–3524

    2023

  48. [56]

    Identity-driven three-player generative adversarial network for synthetic-based face recognition,

    J. N. Kolf, T. Rieber, J. Elliesen, F. Boutros, A. Kuijper, and N. Damer, “Identity-driven three-player generative adversarial network for synthetic-based face recognition,” in CVPR, 2023, pp. 806–816

    2023

  49. [57]

    CemiFace: Center-based semi-hard synthetic face generation for face recognition,

    Z. Sun, S. Song, I. Patras, and G. Tzimiropoulos, “CemiFace: Center-based semi-hard synthetic face generation for face recognition,” NeurIPS, 2024

    2024

  50. [58]

    Hyperface: Generating synthetic face recognition datasets by exploring face embedding hypersphere,

    H. O. Shahreza and S. Marcel, “Hyperface: Generating synthetic face recognition datasets by exploring face embedding hypersphere,” ICLR, 2025

    2025

  51. [59]

    Synthetic face datasets generation via latent space exploration from brownian identity diffu- sion,

    D. Geissb ¨uhler, H. O. Shahreza, and S. Marcel, “Synthetic face datasets generation via latent space exploration from brownian identity diffu- sion,” ICML, 2025

    2025

  52. [60]

    AgeDB: The first manually collected, in-the-wild age database,

    S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia, and S. Zafeiriou, “AgeDB: The first manually collected, in-the-wild age database,” in CVPRW, 2017, pp. 1997–2005. 14

    2017

  53. [61]

    Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments,

    T. Zheng, W. Deng, and J. Hu, “Cross-age LFW: A database for studying cross-age face recognition in unconstrained environments,” arXiv preprint arXiv:1708.08197 , 2017

    2017 arXiv

  54. [62]

    Frontal to profile face verification in the wild,

    S. Sengupta, J. Chen, C. D. Castillo, V . M. Patel, R. Chellappa, and D. W. Jacobs, “Frontal to profile face verification in the wild,” in WACV, 2016, pp. 1–9

    2016

  55. [63]

    Cross-pose LFW: A database for studying cross-pose face recognition in unconstrained environments,

    T. Zheng and W. Deng, “Cross-pose LFW: A database for studying cross-pose face recognition in unconstrained environments,” Beijing University of Posts and Telecommunications, Tech. Rep , 2018

    2018

  56. [64]

    Bisenet: Bilateral segmentation network for real-time semantic segmentation,

    C. Yu, J. Wang, C. Peng, C. Gao, G. Yu, and N. Sang, “Bisenet: Bilateral segmentation network for real-time semantic segmentation,” in ECCV, 2018, pp. 325–341

    2018

  57. [65]

    Magface: A universal representation for face recognition and quality assessment,

    Q. Meng, S. Zhao, Z. Huang, and F. Zhou, “Magface: A universal representation for face recognition and quality assessment,” in CVPR, 2021, pp. 14 225–14 234

    2021

  58. [66]

    How does gender balance in training data affect face recognition accuracy?

    V . Albiero, K. Zhang, and K. W. Bowyer, “How does gender balance in training data affect face recognition accuracy?” in IJCB. IEEE, 2020, pp. 1–10

    2020

  59. [67]

    img2pose: Face alignment and detection via 6dof, face pose estimation,

    V . Albiero, X. Chen, X. Yin, G. Pang, and T. Hassner, “img2pose: Face alignment and detection via 6dof, face pose estimation,” in CVPR, 2021, pp. 7617–7627

    2021

  60. [68]

    A dual-direction attention mixed feature network for facial expression recognition,

    S. Zhang, Y . Zhang, Y . Zhang, Y . Wang, and Z. Song, “A dual-direction attention mixed feature network for facial expression recognition,” Electronics, vol. 12, no. 17, p. 3595, 2023

    2023

  61. [69]

    Affectnet: A database for facial expression, valence, and arousal computing in the wild,

    A. Mollahosseini, B. Hasani, and M. H. Mahoor, “Affectnet: A database for facial expression, valence, and arousal computing in the wild,” IEEE Transactions on Affective Computing , vol. 10, no. 1, pp. 18–31, 2017

    2017

  62. [70]

    Logicnet: A logical consistency embedded face attribute learning network,

    H. Wu, S. Tian, H. Li, and K. W. Bowyer, “Logicnet: A logical consistency embedded face attribute learning network,” WACV, 2025

    2025

  63. [71]

    Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation,

    K. Karkkainen and J. Joo, “Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation,” in WACV, 2021, pp. 1548–1558

    2021

  64. [72]

    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,” in CVPR, 2019, pp. 4690–4699

    2019

  65. [73]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  66. [74]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” in CVPR, 2022, pp. 16 000– 16 009

    2022

  67. [75]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018, pp. 586–595

    2018

  68. [76]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” ICLR, 2015

    2015

  69. [77]

    Toward robust and unconstrained full range of rotation head pose estimation,

    T. Hempel, A. A. Abdelrahman, and A. Al-Hamadi, “Toward robust and unconstrained full range of rotation head pose estimation,” IEEE Transactions on Image Processing , vol. 33, pp. 2377–2387, 2024

    2024

  70. [78]

    Partial FC: training 10 million identities on a single machine,

    X. An, X. Zhu, Y . Gao, Y . Xiao, Y . Zhao, Z. Feng, L. Wu, B. Qin, M. Zhang, D. Zhang, and Y . Fu, “Partial FC: training 10 million identities on a single machine,” in ICCVW, 2021, pp. 1445–1449

    2021

  71. [79]

    Labeled faces in the wild: A database forstudying face recognition in unconstrained environments,

    G. B. Huang, M. Mattar, T. Berg, and E. Learned-Miller, “Labeled faces in the wild: A database forstudying face recognition in unconstrained environments,” in Workshop on faces in’Real-Life’Images: detection, alignment, and recognition , 2008

    2008

  72. [80]

    Morph: A longitudinal image database of normal adult age-progression,

    K. Ricanek and T. Tesafaye, “Morph: A longitudinal image database of normal adult age-progression,” in IEEE F&G, 2006, pp. 341–345

    2006

  73. [81]

    Distinguishing identical twins by face recognition,

    P. J. Phillips, P. J. Flynn, K. W. Bowyer, R. W. V . Bruegge, P. J. Grother, G. W. Quinn, and M. Pruitt, “Distinguishing identical twins by face recognition,” in IEEE F&G, 2011, pp. 185–192

    2011

  74. [82]

    Id 3:identity-preserving-yet-diversified diffusion models for synthetic face recognition,

    S. Li, J. Xu, J. Wu, M. Xiong, A. Deng, J. Ji, Y . Huang, W. Feng, S. Ding, and H. Bryan, “Id 3:identity-preserving-yet-diversified diffusion models for synthetic face recognition,” arXiv preprint arXiv:2409.17576, 2024

    2024 arXiv

  75. [83]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

    2016

  76. [84]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in CVPR, 2018, pp. 7132–7141

    2018

  77. [85]

    Analysis of gender inequality in face recognition accuracy,

    V . Albiero, K. Ks, K. Vangara, K. Zhang, M. C. King, and K. W. Bowyer, “Analysis of gender inequality in face recognition accuracy,” inWACVW, 2020, pp. 81–89

    2020

  78. [86]

    Issues related to face recognition accuracy varying based on race and skin tone,

    K. Krishnapriya, V . Albiero, K. Vangara, M. C. King, and K. W. Bowyer, “Issues related to face recognition accuracy varying based on race and skin tone,” IEEE TTS, vol. 1, no. 1, pp. 8–20, 2020

    2020

  79. [87]

    ‘racist and sexist’ facial recognition cameras could lead to false arrests,

    T. Hoggins, “‘racist and sexist’ facial recognition cameras could lead to false arrests,” Dec. 20 2019, https://www.telegraph.co.uk/technology/2019/12/20/racist-sexist-facial- recognition-cameras-could-lead-false-arrests/

    2019

  80. [88]

    Facial recognition is accurate, if you’re a white guy,

    S. Lohr, “Facial recognition is accurate, if you’re a white guy,” The New York Times, Feb. 9 2018

    2018

  81. [89]

    Emerging from AI utopia,

    E. Santow, “Emerging from AI utopia,” Science, April 2020

    2020

  82. [90]

    Gender and racial bias found in amazon’s facial recognition technology (again),

    J. Vincent, “Gender and racial bias found in amazon’s facial recognition technology (again),” The Verge , Jan. 25 2019, https://www.theverge.com/2019/1/25/18197137/amazon-rekognition- facial-recognition-bias-race-gender. VIII. B IOGRAPHY SECTION Haiyu Wu received the B.S. degr...

    2019

  83. [2015]

    for distinguished contributions to the field of computer vision and pattern recognition, biometrics, object recognition and data science

    He is an associate professor with the School of Computing, Australian National University. He regularly serves as Area Chair for leading confer- ences including ICCV , CVPR, ECCV , NeurIPS and ICML. He was Program Co-Chair for ACM MM 2024 and will be General Co-Chair for A VSS...

    2024

Pith tools

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