Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Foundation versus Domain-specific Models: Performance Comparison, Fusion, and Explainability in Face Recognition

T0 review · 3 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 Domain-specific face recognition models still beat foundation models, and fusing them improves low-error performance.

desk verdict An extensive zero-shot comparison of foundation models vs. dedicated face recognizers, with a promising fusion result that is currently unreproducible because the fusion rule is omitted. read the letter →

arxiv 2507.03541 v2 pith:YDTHNK2F submitted 2025-07-04 cs.CV cs.AI

Redwan Sony , Parisa Farmanifard , Arun Ross , Anil K. Jain This is my paper · ORCID
classification cs.CVcs.AI
keywords facerecognitionfoundationmodelszero-shotscore-levelfusionexplainabilityGPT-4oAdaCLIP
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 asks whether generic foundation models trained on internet images and text can match specialized face-recognition models on the 1:1 face verification task. It argues, with experiments across five benchmarks, that domain-specific models such as AdaFace and ArcFace still outperform zero-shot foundation models at every operating point, especially at very low false-match rates. It then shows that a simple score-level fusion of a foundation model with a face-specific model raises true-match rates at those strict operating points, on IJB-B from 72.64% to 83.31% TMR at 0.0001% FMR for AdaFace plus BLIP2. Finally, it demonstrates that closed-source models like GPT-4o can produce textual explanations for verification decisions and, on selected hard pairs, correct low-confidence errors made by AdaFace.

What carries the argument

The mechanism that carries the argument is complementarity between two image representations: a face-specific embedding trained only on aligned face crops, and a generic vision-language embedding that also encodes surrounding context. The paper pairs these at score level, using similarity scores from AdaFace or ArcFace combined with scores from BLIP2-O-6.7B or CLIP-L-14-336 vision encoders, and shows the combined score recovers true matches at low false-match rates; a second mechanism is prompt-conditioned text generation from closed foundation models, used to provide explanations for verification decisions.

What would settle it

Re-run the IJB-B experiment with a pre-registered fusion rule (for example, unweighted sum of z-normalized scores from AdaFace and BLIP2) fit only on a training split, and check whether the TMR at 0.0001% FMR still exceeds AdaFace's 72.64%; if it does not, the claimed fusion benefit is an artifact of test-set-tuned fusion. A second check: evaluate GPT-4o on a random, independently selected sample of LFW pairs rather than the curated 200-pair set.

Watch

Extended reading notes

Core claim

Domain-specific face recognition models remain clearly superior to zero-shot foundation models on all five benchmark datasets; the performance gap widens as the false-match rate becomes stricter. However, foundation models extract complementary information: they improve when given loosely cropped faces with contextual background, whereas face-specific models degrade on such inputs, and combining the two at score level improves true-match rates at low false-match rates on IJB-B and IJB-C. In addition, GPT-4o under a fully neutral prompt classified all 200 selected LFW pairs correctly and produced feature-level explanations that sometimes identified the right answer when AdaFace was wrong.

Load-bearing premise

The reported fusion improvements assume a score-combination rule and preprocessing that are not specified in the paper; if the weights or normalization were tuned on the same test sets, or if the face crops and alignment differ between the compared models, the fusion gains may not transfer to new data.

Editorial extensions

If this is right

  • At strict false-match rates, zero-shot foundation models alone are not usable for high-security face verification.
  • Loosely cropped inputs help foundation models but hurt face-specific models, so reporting crop size is essential when comparing the two families.
  • Score-level fusion extends the operating range: on IJB-B and IJB-C, the best fused TMR at 0.0001% FMR exceeds either single model.
  • Foundation models can supply post-hoc textual explanations, enabling human review of verification decisions.
  • The complementarity motivates combining face-specific and generic visual features in future FR pipelines.

Reading between the lines

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

  • The context-dependence result implies that foundation-model face performance is systematically underestimated by tight-crop evaluation protocols; a cheap extension would be to add background or context features to FR embeddings instead of using full foundation models.
  • The 200/200 GPT-4o score is on selected pairs and should not be read as general accuracy; a random-sample evaluation would quantify the true explanation and correction rate.
  • If the fusion gain reproduces under a fixed rule, the same complementarity may transfer to other biometrics such as iris or fingerprint, where specialized matchers and generic vision models encode different cues.
  • The observation that GPT-4o followed AdaFace's wrong scores under prompts that mentioned the score, but decided correctly under a fully neutral prompt, suggests anchoring by external scores is a measurable failure mode worth testing explicitly.
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

3 major / 4 minor

Summary. The paper compares domain-specific face recognition (FR) models (AdaFace, ArcFace) with open- and closed-source foundation models on five benchmarks: LFW, WebFace42M-Subset, IJB-B, IJB-C, and IJB-S. The main claims are: (a) zero-shot foundation models are outperformed by FR models on all datasets; (b) foundation models improve when the input is loosely cropped to include context; (c) score-level fusion of a foundation model with an FR model improves true-match rates at low false-match rates; and (d) closed-source models such as GPT-4o and Grok-4 can provide textual explanations and sometimes correct low-confidence AdaFace decisions. The evidence consists of TMR tables, ROC curves, qualitative examples, and prompt-based explainability experiments.

Significance. If the fusion recipe is fully specified and the explainability protocol is tightened, this is a useful empirical contribution: it quantifies the gap between generic foundation models and state-of-the-art FR models on several benchmarks, documents a context-sensitivity effect, and illustrates how foundation models can articulate facial cues. Strengths include the breadth of models evaluated (35 open-source, 3 closed-source), the zero-shot evaluation without fine-tuning, and the explicit reporting of failure cases. The main weaknesses are that the headline fusion gains are not currently reproducible from the manuscript, and the explainability results depend on a selection protocol that is not fully described.

major comments (3)
  1. [Section 4.2] The fusion rule is not specified. The paper reports TMR improvements such as AdaFace+BLIP2-O-6.7B on IJB-B rising from 72.64% to 83.31% at FMR=0.0001% without giving the score normalization, combination formula, fusion weight, or the procedure for setting FMR thresholds on the fused scores. Because AdaFace cosine similarity and BLIP2 alignment scores are not directly comparable, the reader cannot determine whether the gain reflects complementary information or ad hoc calibration and threshold selection. Please specify the exact preprocessing (cropping/alignment per model), score normalization, fusion equation, and threshold estimation protocol, and report baselines such as max-score, equal-weight sum of rank-normalized scores, and simple averaging, together with confidence intervals. This is load-bearing for claim (c).
  2. [Section 3 / Table 2] The reported pair counts for WebFace42M-Subset are internally inconsistent. Table 2 lists 50,000 subjects and 250,000 images, but the genuine-pair count (100,000) and impostor-pair count (1,249,875,000) are exactly those of 10,000 subjects with 5 images each (50,000 images). The text states 'five disjoint sets of 10K identities, each with 5 images (50K total)', which also suggests per-split counts rather than totals. Please clarify whether Table 2 reports totals or per-split counts and correct the numbers accordingly; this affects the interpretation of the large-scale benchmark claim and the evaluation protocol in Table 4.
  3. [Section 4.3 / Tables 6 and 8] The selection of the 200 LFW pairs and the 4,000 IJB-S hard pairs is not described sufficiently. Table 6 reports 200/200 correct decisions under Prompt-3, but the reader does not know how the 50 FNM, 50 FM, 50 TNM, and 50 TM pairs were sampled from AdaFace's error sets, whether any filtering by image quality or detection failure was applied, or how API non-determinism was handled. Without a fixed random sampling protocol and full confusion matrices, the perfect score is not interpretable and the 'resolving low-confidence decisions' claim is vulnerable to selection bias. In addition, the text states that GPT-4o achieved 78.80% overall accuracy on the IJB-S hard set, while Table 8 reports 74.45%; please reconcile this discrepancy.
minor comments (4)
  1. [Section 4.1] The reference assignment is incorrect: the sentence 'IJB-B [50] and IJB-C [21], IJB-S (Still to Still) [31]' should be IJB-C [31] and IJB-S [21].
  2. [Abstract / Section 4.3] The abstract and conclusion state that score-level fusion improves accuracy, but Section 4.3 reports that score-level fusion with GPT-4o and Grok-4 on the broader IJB-S set 'did not lead to any improvement in verification performance.' Please qualify claim (c) to the open-source fusion experiments or explain the difference.
  3. [Section 4.1, text near Table 3] The text says 'Both AdaFace and ArcFace models produced True Match Rates (TMRs) above 99.09%' at 112x112 and 160x160, but Table 3 reports ArcFace at 98.86% for 112x112. This should be corrected to avoid overstating the result.
  4. [Section 4, introductory paragraph] The paper reports only a subset of foundation models 'demonstrating performance close to the FR models.' Please define how 'close' was determined, or state explicitly that the selection was based on the supplementary results, so that the reader can assess potential selection bias in the main-text tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's comparisons, zero-shot evaluations, and explainability checks are self-contained; the underspecified fusion rule and a dataset-table inconsistency are correctness/reproducibility issues, not circularity.

full rationale

This is an empirical benchmarking paper, not a derivation chain: reported quantities (TMR at FMR for each model, fusion scores, and GPT-4o/Grok-4 classification accuracies) are measured on held-out benchmarks using fixed pretrained weights rather than derived from fitted equations. No parameter is fitted to test labels and then renamed as a prediction. The fusion section (Section 4.2) is underspecified—the text only says 'we perform score-level fusion as shown in Figure 2' without giving the fusion rule, normalization, or weight selection—but underspecification is a reproducibility/correctness problem, not circularity; nothing in the paper indicates the fusion weights or thresholds were optimized on the test impostor distribution. The explainability experiments select hard LFW and IJB-S pairs using AdaFace scores (Section 4.3: 'chosen from AdaFace’s correctly and incorrectly classified pairs' and 'challenging genuine and impostor pairs identified by AdaFace'), but the evaluation prompt (Prompt-3) does not include those scores, so the GPT-4o/Grok-4 classifications are not folded-back inputs; sample selection affects difficulty but does not define the measured outcome. The authors do cite their own prior benchmark, Sony et al. [45], for the model-selection protocol ('following the approach in [45]'), but this self-citation is not load-bearing: the models, datasets, and evaluation are described independently, and no central claim reduces to a uniqueness theorem or derived formula from [45]. One internal inconsistency exists—Table 2 gives WebFace42M-Subset as 50,000 subjects/250,000 images, while its stated pair counts (100,000 genuine and 1,249,875,000 impostor) correspond to 10,000 subjects with 5 images each—but this is a dataset-reporting error, not a circular step. No equation or result is equivalent to its own input by construction, so the appropriate circularity score is 0.

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

The paper introduces no new entities and no fitted parameters; its claims come from empirical evaluation of existing models on standard benchmarks. The main unstated premises are the fidelity of the benchmark protocols, the unspecified preprocessing/fusion pipeline, and the stability of closed-source APIs.

assumptions (3)
  • domain assumption Standard benchmark protocols (LFW exhaustive, WebFace42M-Subset, IJB-B, IJB-C, IJB-S) are correctly implemented and representative of face recognition performance.
    The paper's evaluations rely on published protocols and ground-truth labels; any protocol implementation error would affect all reported TMRs.
  • domain assumption The exact cropping and alignment pipeline for LFW at 112x112, 160x160, and 250x250 does not introduce artifacts that change model rankings.
    Section 4.1 compares models across crop sizes without specifying the preprocessing code, so the comparison assumes a consistent and sensible pipeline.
  • domain assumption Closed-source models (GPT-4o, Gemini, Grok-4) are assumed to behave consistently at the time of evaluation, and their API outputs are taken as valid evidence.
    Sections 4.3 and 5 describe API-based experiments with no version pinning; results may not hold for updated model versions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Foundation versus Domain-specific Models: Performance Comparison, Fusion, and Explainability in Face Recognition." pith.science (2026). https://pith.science/paper/YDTHNK2F

@misc{pith2026250703541,
  author       = {Pith},
  title        = {Pith review of: Foundation versus Domain-specific Models: Performance Comparison, Fusion, and Explainability in Face Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDTHNK2F}},
  note         = {Machine review of arXiv:2507.03541}
}
read the original abstract

In this paper, we address the following question: How do generic foundation models (e.g., CLIP, BLIP, GPT-4o, Grok-4) compare against a domain-specific face recognition model (viz., AdaFace or ArcFace) on the face recognition task? Through a series of experiments involving several foundation models and benchmark datasets, we report the following findings: (a) In all face benchmark datasets considered, domain-specific models outperformed zero-shot foundation models. (b) The performance of zero-shot generic foundation models improved on over-segmented face images compared to tightly cropped faces, thereby suggesting the importance of contextual clues. (c) A simple score-level fusion of a foundation model with a domain-specific face recognition model improved the accuracy at low false match rates. (d) Foundation models, such as GPT-4o and Grok-4, are able to provide explainability to the face recognition pipeline. In some instances, foundation models are even able to resolve low-confidence decisions made by AdaFace, thereby reiterating the importance of combining domain-specific face recognition models with generic foundation models in a judicious manner.

Figures

Figures reproduced from arXiv: 2507.03541 by the authors.

Figure 1
Figure 1. Example face pairs from IJB-B [50] and IJB-S [21] datasets misclassified by AdaFace [22] but correctly classified after the fusion of AdaFace and BLIP2 [27] or AdaFace and GPT-4o [36]. Abstract In this paper, we address the following question: How do generic foundation models (e.g., CLIP, BLIP, GPT-4o, Grok-4) compare against a domain-specific face recogni￾tion model (viz., AdaFace or ArcFace) on the face recogni￾ti… view at source ↗
Figure 2
Figure 2. Fusion of face recognition model with foundation model [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Example images from the LFW dataset, showing face [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: ROC curves for the fusion of foundation model BLIP2-O-6.7B with AdaFace and ArcFace on the (a) IJB-B, (b) IJB-C, and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: GPT-4o correctly classified these “edge” cases from the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ArtFace: Towards Historical Portrait Face Identification via Model Adaptation

    cs.CV 2025-08 conditional novelty 4.0 of 10

    Fine-tuning CLIP with LoRA and fusing its embeddings with an adapted face recognition network reduces equal error rate on the Historical Faces dataset to 9.9%.

Reference graph

Works this paper leans on

55 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    Technical report, National Institute of Standards and Technology, 2025

    Face Recognition Technology Evaluation (FRTE) 1:1 Verifi- cation. Technical report, National Institute of Standards and Technology, 2025. Report published 2025-06-25. 1

    work page 2025

  2. [2]

    Face Recognition in the age of CLIP & Billion image datasets

    Aaditya Bhat and Shrey Jain. Face Recognition in the Age of CLIP & Billion Image Datasets. arXiv preprint arXiv:2301.07315, 2023. 2

    work page Pith review arXiv 2023

  3. [3]

    On the Opportunities and Risks of Foundation Models.arXiv preprint arXiv:2108.07258, 2021

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Alt- man, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the Opportunities and Risks of Foundation Models.arXiv preprint arXiv:2108.07258, 2021. 2

    arXiv 2021

  4. [4]

    Emerg- ing Properties in Self-Supervised Vision Transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing Properties in Self-Supervised Vision Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650–9660, 2021. 2, 3

    work page 2021

  5. [5]

    InternVL: Scaling Up Vision Foundation Models and Aligning for Generic Visual-linguistic Tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling Up Vision Foundation Models and Aligning for Generic Visual-linguistic Tasks. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 24185–24198,

    work page

  6. [6]

    Reproducible Scal- ing Laws for Contrastive Language-Image Learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh- mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible Scal- ing Laws for Contrastive Language-Image Learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2818–2829, 2023. 2, 3, 4

    work page 2023

  7. [7]

    FRounda- tion: Are Foundation Models Ready for Face Recognition? Image and Vision Computing, 132:104815, 2025

    Tahar Chettaoui, Naser Damer, and Fadi Boutros. FRounda- tion: Are Foundation Models Ready for Face Recognition? Image and Vision Computing, 132:104815, 2025. 2

    work page 2025

  8. [8]

    MS1MV2 Dataset

    InsightFace Contributors. MS1MV2 Dataset. https: //github.com/deepinsight/insightface . Ac- cessed: 2025-06-21. 3

    work page 2025

Show all 55 references
  1. [9]

    DeepSeekMoE: Towards Ultimate Ex- pert Specialization in Mixture-of-Experts Language Models

    Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Yu Wu, et al. DeepSeekMoE: Towards Ultimate Ex- pert Specialization in Mixture-of-Experts Language Models. arXiv preprint arXiv:2401.06066, 2024. 2

    2024 arXiv

  2. [10]

    How Good is ChatGPT at Face Biomet- rics? A First Look Into Recognition, Soft Biometrics, and Explainability

    Ivan Deandres-Tame, Ruben Tolosana, Ruben Vera- Rodriguez, Aythami Morales, Julian Fierrez, and Javier Ortega-Garcia. How Good is ChatGPT at Face Biomet- rics? A First Look Into Recognition, Soft Biometrics, and Explainability. IEEE Access, 2024. 2, 5

    2024

  3. [11]

    ArcFace: Additive Angular Margin Loss for Deep Face Recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: Additive Angular Margin Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pages 4690–4699, 2019. 1, 3, 4, 5

    2019

  4. [12]

    Lightweight Face Recognition Chal- lenge

    Jiankang Deng, Jia Guo, Debing Zhang, Yafeng Deng, Xi- angju Lu, and Song Shi. Lightweight Face Recognition Chal- lenge. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision Workshops (ICCVW), pages 0–0,

  5. [13]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition ...

    2021

  6. [14]

    ChatGPT Meets Iris Bio- metrics

    Parisa Farmanifard and Arun Ross. ChatGPT Meets Iris Bio- metrics. In IEEE International Joint Conference on Biomet- rics (IJCB), pages 1–10, 2024. 2

    2024

  7. [15]

    Iris-SAM: Iris Segmenta- tion Using a Foundation Model

    Parisa Farmanifard and Arun Ross. Iris-SAM: Iris Segmenta- tion Using a Foundation Model. In International Conference on Pattern Recognition and Artificial Intelligence (ICPRAI), pages 394–409, 2024. 2

    2024

  8. [16]

    ChatGPT and Biometrics: An Assessment of Face Recognition, Gender Detection, and Age Estimation Capabilities

    Ahmad Hassanpour, Yasamin Kowsari, Hatef Otroshi Shahreza, Bian Yang, and S ´ebastien Marcel. ChatGPT and Biometrics: An Assessment of Face Recognition, Gender Detection, and Age Estimation Capabilities. In IEEE In- ternational Conference on Image Processing (ICIP) , pages 322...

    2024

  9. [17]

    Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller

    Gary B. Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environ- ments. In Workshop on Faces in ‘Real-Life’ Images: De- tection, Alignment, and Recognition, 2008. 4

    2008

  10. [18]

    OpenCLIP: Open-Source CLIP Reproduction

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Ali Farhadi, Gordon Wetzstein, Benjamin Recht, and Afshin Rostamizadeh. OpenCLIP: Open-Source CLIP Reproduction. In NeurIPS Datasets and Benchmarks Track, 2021. 3

    2021

  11. [19]

    Jain, Arun A

    Anil K. Jain, Arun A. Ross, Karthik Nandakumar, and Thomas Swearingen. Introduction to Biometrics. Springer, 2nd edition, 2025. 1

    2025

  12. [20]

    Scaling Up Visual and Vision-language Represen- tation Learning With Noisy Text Supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-language Represen- tation Learning With Noisy Text Supervision. In Inter- national Conference on Machine Learning (ICML) , page...

    2021

  13. [21]

    Kalka, Brianna Maze, James A

    Nathan D. Kalka, Brianna Maze, James A. Duncan, Kevin O’Connor, Stephen Elliott, Kaleb Hebert, Julia Bryan, and Anil K. Jain. IJB–S: IARPA Janus Surveillance Video Benchmark. In Proceedings of the IEEE 9th International Conference on Biometrics Theory, Applications and Systems...

    2018

  14. [22]

    Jain, and Xiaoming Liu

    Minchul Kim, Anil K. Jain, and Xiaoming Liu. AdaFace: Quality Adaptive Margin for Face Recognition. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 4, 5, 7

    2022

  15. [23]

    Jain, and Xiaoming Liu

    Minchul Kim, Anil K. Jain, and Xiaoming Liu. 50 Years of Automated Face Recognition. arXiv preprint arXiv:2505.24247, 2025. 1

    2025

  16. [24]

    Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment Anything. In Proceedings of the IEEE/CVF International Conference on Computer Visio...

    2023

  17. [25]

    Exploring ChatGPT for Face Presentation Attack Detection in Zero and Few-Shot in-Context Learning

    Alain Komaty, Hatef Otroshi Shahreza, Anjith George, and Sebastien Marcel. Exploring ChatGPT for Face Presentation Attack Detection in Zero and Few-Shot in-Context Learning. In IEEE/CVF Winter Conference on Applications of Com- puter Vision Workshops (WACVW), pages 1602–1611, 2025. 2

    2025

  18. [26]

    BLIP: Bootstrapping Language-image Pre-training for Uni- fied Vision-language Understanding and Generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping Language-image Pre-training for Uni- fied Vision-language Understanding and Generation. In In- ternational Conference on Machine Learning (ICML), pages 12888–12900, 2022. 2, 3, 5

    2022

  19. [27]

    BLIP-2: Bootstrapping Language-image Pre-training With Frozen Image Encoders and Large Language Models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping Language-image Pre-training With Frozen Image Encoders and Large Language Models. In In- ternational Conference on Machine Learning (ICML), pages 19730–19742, 2023. 1, 2, 4

    2023

  20. [28]

    Improved Baselines With Visual Instruction Tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines With Visual Instruction Tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26296–26306, 2024. 2, 4

    2024

  21. [29]

    LLaV A-NeXT: Im- proved Reasoning, OCR, and World Knowledge, 2024

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaV A-NeXT: Im- proved Reasoning, OCR, and World Knowledge, 2024. 2, 3

    2024

  22. [30]

    SphereFace: Deep Hypersphere Embed- ding for Face Recognition

    Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. SphereFace: Deep Hypersphere Embed- ding for Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 212–220, 2017. 1

    2017

  23. [31]

    Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K

    Brianna Maze, Jocelyn Adams, James A. Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K. Jain, W. Tyler Niggel, Janet Anderson, Jordan Cheney, et al. IARPA Janus Benchmark-C: Face Dataset and Protocol. In International Conference on Biometrics (ICB), pages 158–165, 2018. 3, 4

    2018

  24. [32]

    Segment Anything Model for Medical Image Analysis: An Experimental Study

    Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment Anything Model for Medical Image Analysis: An Experimental Study. Medical Image Analysis, 89:102918, 2023. 2

    2023

  25. [33]

    MagFace: A Universal Representation for Face Recognition and Quality Assessment

    Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. MagFace: A Universal Representation for Face Recognition and Quality Assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14225–14234, 2021. 1

    2021

  26. [34]

    Kartik Narayan, VS Vibashan, and Vishal M. Patel. FaceXBench: Evaluating Multimodal LLMs on Face Under- standing. CoRR, abs/2501.10360, 2025. arXiv preprint. 2

    2025

  27. [35]

    GPT-4 Technical Report

    OpenAI. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774, 2023. 2, 5

    2023 arXiv

  28. [36]

    GPT-4o System Card

    OpenAI. GPT-4o System Card. https://openai.com/ index/gpt- 4o- system- card/ , 2024. Accessed: 2025-05-11. 1, 2, 3, 5, 7

    2024

  29. [37]

    DINOv2: Learning Robust Visual Features Without Super- vision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features Without Super- vision. arXiv preprint arXiv:2304.07193, 2023. 3

    2023 arXiv

  30. [38]

    FoundPAD: Foundation Models Reloaded for Face Presen- tation Attack Detection, 2025

    Guray Ozgur, Eduarda Caldeira, Tahar Chettaoui, Fadi Boutros, Raghavendra Ramachandra, and Naser Damer. FoundPAD: Foundation Models Reloaded for Face Presen- tation Attack Detection, 2025. 2

    2025

  31. [39]

    Kosmos-2: Grounding Multimodal Large Language Models to the World

    Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding Multimodal Large Language Models to the World. arXiv preprint arXiv:2306.14824, 2023. 2, 3

    2023 arXiv

  32. [40]

    Learning Transferable Visual Models from Natural Language Super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language Super- vision. In Proceedings of the International Conference on...

    2021

  33. [41]

    SAM 2: Segment Anything in Images and Videos

    Nikhila Ravi, Valentin Gabeur, and Others. SAM 2: Segment Anything in Images and Videos. In ICLR, 2025. 2

    2025

  34. [42]

    A System- atic Survey of Prompt Engineering in Large Language Models: Techniques and Applications

    Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A System- atic Survey of Prompt Engineering in Large Language Models: Techniques and Applications. arXiv preprint arXiv:2402.07927, 2024. 6

    2024 arXiv

  35. [43]

    FaceNet: A Unified Embedding for Face Recognition and Clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 815–823, 2015. 1

    2015

  36. [44]

    Foundation Models and Biometrics: A Survey and Outlook

    Hatef Otroshi Shahreza and S ´ebastien Marcel. Foundation Models and Biometrics: A Survey and Outlook. TechRxiv,

  37. [45]

    Benchmarking Foundation Models for Zero-Shot Biometric Tasks

    Redwan Sony, Parisa Farmanifard, Hamzeh Alzwairy, Ni- tish Shukla, and Arun Ross. Benchmarking Foundation Models for Zero-Shot Biometric Tasks. arXiv preprint arXiv:2505.24214, 2025. 2, 3

    2025 arXiv

  38. [46]

    Towards Iris Presentation Attack Detection with Foundation Models

    Juan E Tapia, L ´azaro Janier Gonz ´alez-Soler, and Christoph Busch. Towards Iris Presentation Attack Detection with Foundation Models. arXiv preprint arXiv:2501.06312, 2025. 2

    2025 arXiv

  39. [47]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, et al. Gemini: A Family of Highly Capable Multimodal Models. arXiv preprint arXiv:2312.11805, 2025. 2, 3, 5

    2025 arXiv

  40. [48]

    CosFace: Large Margin Cosine Loss for Deep Face Recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large Margin Cosine Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 5265–5274,

  41. [49]

    DeepSeek-VL: Towards Scaling Up Vision-Language Mod- els

    Zikang Wang, Ziyang Li, Chong He, Tang Shen, Xiaoyu Zhang, Zihang Liu, Haoyuan Tang, Feng Wang, et al. DeepSeek-VL: Towards Scaling Up Vision-Language Mod- els. arXiv preprint arXiv:2401.00001, 2024. 2

    2024 arXiv

  42. [50]

    Jain, James A

    Cameron Whitelam, Emma Taborsky, Austin Blanton, Bri- anna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K. Jain, James A. Duncan, Kristen Allen, Jordan Ch- eney, and Patrick Grother. IARPA Janus Benchmark-B Face Dataset. In Proceedings of the IEEE/CVF Confer- ence on Co...

    2017

  43. [51]

    DeepSeek-VL2: Mixture-of- Experts Vision-Language Models for Advanced Multimodal Understanding

    Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. DeepSeek-VL2: Mixture-of- Experts Vision-Language Models for Advanced Multimodal Understanding. arXiv preprint arXiv:2412.10302, 2024. 2, 3

    2024 arXiv

  44. [52]

    Grok-4 Language Model

    xAI. Grok-4 Language Model. https://x.ai/news/ grok-4, 2025. Accessed: July 29, 2025. 2, 3, 5, 7

    2025

  45. [53]

    Cross-Pose LFW: A Database for Studying Cross-Pose Face Recognition in Un- constrained Environments

    Tong Zheng and Weihong Deng. Cross-Pose LFW: A Database for Studying Cross-Pose Face Recognition in Un- constrained Environments. Technical Report 18-01, Beijing University of Posts and Telecommunications, 2018. 3, 4, 7

    2018

  46. [54]

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Mod- els, (Technical Report)

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shen- glong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Mod- els, (Technical Report). arXiv preprint arXiv:2504.10479 ,

    arXiv

  47. [55]

    WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition

    Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Da- long Du, et al. WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and...

    2021

Pith tools

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