Pith. sign in

REVIEW 4 major objections 6 minor 81 references

GIF: Generative Inspiration for Face Recognition at Scale

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Replacing scalar labels with structured integer codes makes face-recognition training cost logarithmic in the number of identities and improves accuracy on IJB-B and IJB-C.

desk verdict The tokenized-label trick genuinely makes classifier cost logarithmic and mostly improves accuracy, but the paper overclaims on its own ViT numbers and never verifies that the CLIP-derived 'structure' in the codes is what drives the gains. read the letter →

arxiv 2505.03012 v1 pith:4MN6B5GK submitted 2025-05-05 cs.CV

classification cs.CV
keywords facerecognitionstructuredidentitycodesefficienttraininglogarithmicscalinghierarchicalk-meanshypersphericaluniformityminoritycollapsegenerativeretrieval
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 sets out to remove the main scaling bottleneck in large-scale face-recognition training: the softmax classifier, whose cost grows linearly with the number of identities. It claims that atomic scalar labels can be replaced by structured identity codes, short sequences of integer tokens, and that a face encoder trained to predict the tokens instead of a class label reduces label-matching cost to $O(\log m)$ while actually improving verification accuracy. The identity codes are built by starting from per-identity semantic visual embeddings, spreading those embedding points uniformly over the unit hypersphere, and then applying hierarchical $k$-means to route every identity to a code path. On the IJB-B and IJB-C benchmarks the method reports gains over prior efficient-training and standard training baselines at TAR@FAR=$10^{-4}$, with larger gains on the more imbalanced datasets. If the claim holds, million-identity face training becomes feasible on much smaller GPU budgets.

What carries the argument

The load-bearing object is the identity-code tree. A structured identity code is a sequence of integers that names a path through a tree: the root branches into $v$ clusters, each cluster branches again, and the leaf path is the code. To build that tree the paper maps every identity to a point on the unit hypersphere, using the mean of semantic visual embeddings, then applies a Gaussian-potential uniformity loss to spread the $m$ points as evenly as possible, and finally runs hierarchical $k$-means on the spread points. During training, $l$ small classifiers each predict one token of the code, with each token normalized over only $v$ possibilities, plus a regression loss that pulls the face embedding toward its assigned point. The mechanism is what converts the $O(m)$ normalization of angular-margin softmax into $l$ normalizations of size $v$. The code structure carries the semantics: identities that share a token are intended to be generically similar.

What would settle it

A concrete check is to measure whether the uniformity optimization preserves semantic structure: compare the average semantic-embedding similarity of identity pairs that share a first code token against pairs that do not, using the optimized code tree. If the first-token groups are no more similar than random groups, the structured part of the code is not carrying face semantics. A second check is to keep the same tree shape but randomly permute which identity lands in which leaf; the paper's atomic-code ablation predicts this collapse, so it would isolate hierarchy from uniformity.

Watch

Extended reading notes

Core claim

The paper's central claim is that face-recognition training need not match a feature vector against one centroid per identity. Each identity is assigned a structured code $c_{y}=(c_{y}^{1},\dots,c_{y}^{l})$ with every token in $[0,v-1]$, and training solves $l$ parallel $v$-way classification problems instead of one $m$-way problem; since $v^{l}\ge m$, the classifier cost scales as $O(\log m)$. The code is not arbitrary: per-identity reference points are initialized from semantic visual embeddings, optimized to be uniformly spread on the unit hypersphere with a Gaussian-potential uniformity loss, and then partitioned by hierarchical $k$-means so that similar identities share token prefixes. A regression term pulls the embedding toward its assigned reference point, and the combined objective is claimed to avoid 'minority collapse' because the reference-point layout is independent of how many images each identity has. The paper reports that this scheme outperforms prior efficient-training methods, with gains such as 1.52% and 0.6% at TAR@FAR=$10^{-4}$ on IJB-B and IJB-C, while also beating standard distributed-training baselines.

Load-bearing premise

The whole scheme rests on the assumption that spreading the reference points uniformly over the sphere preserves enough of their original semantic ordering for the hierarchical codes to group genuinely similar faces; if that ordering is destroyed, shared code tokens become meaningless and the method reduces to atomic codes, which the paper's own ablation shows fail.

Editorial extensions

If this is right

  • With the classifier head no longer proportional to identity count, training on datasets with tens of millions of identities fits on far fewer GPUs.
  • The reported gains on IJB-B and IJB-C, particularly on imbalanced WebFace subsets, indicate that dropping the global softmax normalization does not cost accuracy and can improve it.
  • Because the reference-point layout is fixed before training and does not depend on per-identity sample counts, the minority-collapse failure mode of angular-margin softmax is removed by construction.
  • The embedding produced by the backbone remains the deployment representation, so the method changes training cost without changing verification or identification protocols.
  • Structured codes are necessary: the paper's ablations show replacing them with random atomic codes drops performance to near chance.

Reading between the lines

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

  • Editorial inference: the claimed $O(\log m)$ saving is a per-training-iteration statement about the classifier head; the one-time tokenization still processes all $m$ identities, so total pipeline cost remains at least linear in $m$ for a fixed dataset.
  • Editorial inference: if the hierarchical code tree genuinely preserves semantic neighborhoods after the uniformity step, the tree itself could be reused for coarse-to-fine retrieval or for organizing synthetic identities, applications the paper does not explore.
  • Editorial inference: the dependence on a pretrained semantic encoder means the method's transfer to other recognition domains, such as vehicle or product recognition, is a testable extension; the paper only demonstrates faces.
  • Editorial inference: the reported memory plateau suggests that at very large identity counts the dominant cost may shift from the classifier to the backbone and feature pipeline, so the practical ceiling is no longer the label space.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes GIF, a face-recognition training scheme that replaces scalar identity labels with structured integer codes. Code vectors are initialized from per-identity CLIP embeddings, optimized for hyperspherical uniformity via a Gaussian-potential loss, and then converted to identity codes by hierarchical k-means. Training predicts the l tokens of each identity code through l parallel v-way classifiers, augmented by an angular-regression loss to the assigned code vector, yielding a claimed O(log m) training cost instead of the O(m) cost of a conventional softmax classifier. Experiments on MS1MV2, WebFace4M, WebFace12M, and WebFace42M with ResNet-100 and ViT-B backbones report improvements over prior efficient-training methods on IJB-B and IJB-C, along with memory and speed advantages.

Significance. If the empirical claims hold, GIF is a useful contribution to large-scale face recognition: it changes the scaling of classifier cost and memory from linear to logarithmic in the number of identities, while reporting consistent gains on IJB-B and IJB-C across multiple training-set scales. The paper also provides a constructive ablation structure—loss components, structured versus atomic codes, code length, token range, and CLIP-versus-DINO initialization—and promises public code. The main open question is whether the learned codes are genuinely structured in the semantic sense claimed, and whether that structure is responsible for the accuracy gains, rather than the uniformity-regularized prototype regression alone.

major comments (4)
  1. [§4.5.2] The statement that 'initializing code vectors from a random distribution instead of using the CLIP visual encoder is the same as having atomic codes' is asserted but not tested. A random-initialization control with the same Gaussian-potential optimization and hierarchical k-means is missing. If that control retains the reported accuracy, the semantic structure of the codes is not load-bearing for the gains; if it fails, the paper still needs to demonstrate directly that the optimized code vectors preserve CLIP-derived semantic neighborhoods. This experiment is necessary to support the central contribution claim of structured identity codes.
  2. [§3.4.1 / Table 3] Table 3 shows that CLIP initialization without the uniformity optimization collapses to 24.30% TAR@FAR=1e-4 on IJB-B, and Figure 5d shows that atomic codes fail, so each component is claimed necessary. However, the paper does not verify that the Gaussian-potential optimization preserves the semantic ordering of the CLIP-initialized code vectors. If the repulsive optimization substantially reorders the vectors, the hierarchical k-means codes no longer reflect the 'similar generic information' that the structured-code claim relies on. I request a quantitative check, such as k-nearest-neighbor overlap or normalized mutual information between the initial and optimized code vectors, or an evaluation with a randomly permuted initial H followed by the same optimization pipeline.
  3. [Table 1] The WebFace42M/ViT row reports AgeDB=96.16 for GIF versus 98.53 for PFC, a drop of 2.37 percentage points, which contradicts the text in §4.4 that 'GIF enhances all previous ET methods across LFW, CFP-FP and AgeDB.' This inconsistency must be resolved: either the table entry is incorrect, or the claim needs to be qualified. Because the abstract and conclusion emphasize universal improvement, this discrepancy is material to the paper's main empirical claim.
  4. [§4.2 / Eq. (7)] Several hyperparameters that directly control the method are not reported. The logit scale gamma in Eq. (7) is introduced in Algorithm 1 but its value is never given; the Gaussian-potential width t in Eq. (4) is not specified; and the exact code length l and token range v per dataset are only described by the range 5 <= v <= 20. Since the O(log m) cost claim depends on the actual l and v choices, and since the empirical results cannot be reproduced without these values, the experimental section needs a full hyperparameter table, including per-dataset settings.
minor comments (6)
  1. [§3.4.1] The phrase 'tow steps tokenization scheme' appears to be a typo for 'two-step tokenization scheme.'
  2. [Table 3 caption] The caption contains 'Abblation,' which should read 'Ablation.'
  3. [Eq. (6)] The notation gi,j is used without defining the matrix G; please clarify that gi,j = G_t(h_i, h_j) with G_t as in Eq. (4).
  4. [§3.4.2] The description of assigning 'an arbitrary number from [0,v-1]' to the l-th token should clarify whether this assignment is fixed before training and whether the resulting leaf codes are unique.
  5. [Figure 5e/f] The caption states that the average performance across five benchmarks is reported, but the figure does not show individual benchmark values or error bars; please make the aggregation and the axis ranges explicit.
  6. [Algorithm 1] Line 25 writes 'cyi = [H_phi_1(z),...,H_phi_l(z)]', but H_phi_j are projection heads, not identity codes; the returned values should be described as predicted token logits or probabilities, not codes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the logarithmic-cost claim follows by construction, and the accuracy gains are established by external evaluation and ablations rather than by fitting the reported numbers.

full rationale

The paper's central claims are (i) replacing an m-way softmax with l parallel v-way classifiers changes classifier cost from O(m) to O(log m), and (ii) structured identity codes derived from CLIP-initialized, uniformity-optimized prototypes plus hierarchical k-means improve verification accuracy over atomic scalar labels. Neither claim reduces to its inputs by construction. The cost claim follows from counting: with code length l and token range v, the code space holds v^l identities, so l ~ log_v(m) and the classifier cost is l*v, giving O(log m); this is an architectural design property, not a number fitted to the evaluation data. The accuracy claim is supported by external evaluation on LFW, CFP-FP, AgeDB, IJB-B, and IJB-C, and by ablations showing that atomic codes fail (Figure 5d), that CLIP-initialized prototypes without uniformity collapse (Table 3), and that the optimized prototype distribution works. The CLIP/DINO replacement experiment (Table 4 in the supplement) further shows the method is not tied to a single pretrained model. The paper does contain an unverified assertion that Gaussian-potential optimization preserves the semantic neighborhoods of the CLIP initialization; that is a missing-control or robustness concern, not a circularity, because the claimed semantic structure is not used to define the reported benchmark outcomes. Hyperparameters such as l and v are chosen through ablations on the same benchmarks, which is a model-selection concern rather than a fitted-input-called-prediction circularity: the final evaluation numbers are not produced by minimizing a training loss on the test outputs. Self-citations, e.g., [50], are used only for background on unbalanced distributions and are not load-bearing. There is no imported uniqueness theorem from the authors' own prior work and no ansatz smuggled in solely via self-citation. Overall, the derivation chain is self-contained against external benchmarks, and no specific equation or fitted parameter is equivalent to the reported predictions by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an external teacher (CLIP/DINO), a geometric uniformity step, and a heuristic clustering step. No new physical or mathematical entities are introduced. The main uncharged assumptions are that the code vectors retain semantics after uniformity optimization and that the claimed logarithmic scaling is realized by the chosen l/v schedule.

free parameters (6)
  • code length l = chosen per dataset so 5<=v<=20; exact values not enumerated
    Controls the capacity of the code vocabulary and the cost l*v; selected from ablations in Fig 5e/f.
  • token range v = 5 to 25, with v<=25 when l=4
    Branching factor of hierarchical k-means; chosen jointly with l to keep the token range small.
  • loss balance gamma (Eq 10) = 1.0
    Weight balancing LC and LAR; set by ablation in Fig 5c.
  • token-weight lambda_j = reported as 1 in Sec 4.2 and as 1/l in Supp Sec 7
    The paper contains both values; the inconsistency is not resolved.
  • logit scale gamma in Eq 7 = not reported
    Used in the v-way softmax outputs; absent from the implementation details.
  • Gaussian potential width t = not reported
    Hyperparameter of Eq 4/6; no value is given in Sec 4.2 or Algorithm 1.
assumptions (4)
  • domain assumption CLIP (or DINO) visual embeddings provide a semantically valid initialization for identity code vectors
    Used in Eq 3 and Sec 3.4.1. Ablations show CLIP-only H performs poorly, so the choice of teacher carries substantial weight.
  • domain assumption Uniformly distributed code vectors on the unit hypersphere maximize inter-identity separation and enable metric-space exploitation
    Borrowed from [14,23,40,64,68,69]; not independently demonstrated for FR beyond Table 3.
  • ad hoc to paper Hierarchical k-means on the optimized code vectors yields codes whose shared prefixes reflect shared generic information
    Stated in Sec 3.4.2 but never directly measured; 'similar generic information' is not defined or validated.
  • domain assumption The face backbone can learn to predict fixed code tokens while regressing to fixed code vectors, and this transfers to open-set verification
    The core training hypothesis; supported by experiments but no convergence or generalization analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GIF: Generative Inspiration for Face Recognition at Scale." pith.science (2026). https://pith.science/paper/4MN6B5GK

@misc{pith2026250503012,
  author       = {Pith},
  title        = {Pith review of: GIF: Generative Inspiration for Face Recognition at Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4MN6B5GK}},
  note         = {Machine review of arXiv:2505.03012}
}
abstract

Aiming to reduce the computational cost of Softmax in massive label space of Face Recognition (FR) benchmarks, recent studies estimate the output using a subset of identities. Although promising, the association between the computation cost and the number of identities in the dataset remains linear only with a reduced ratio. A shared characteristic among available FR methods is the employment of atomic scalar labels during training. Consequently, the input to label matching is through a dot product between the feature vector of the input and the Softmax centroids. Inspired by generative modeling, we present a simple yet effective method that substitutes scalar labels with structured identity code, i.e., a sequence of integers. Specifically, we propose a tokenization scheme that transforms atomic scalar labels into structured identity codes. Then, we train an FR backbone to predict the code for each input instead of its scalar label. As a result, the associated computational cost becomes logarithmic w.r.t. number of identities. We demonstrate the benefits of the proposed method by conducting experiments. In particular, our method outperforms its competitors by 1.52%, and 0.6% at TAR@FAR$=1e-4$ on IJB-B and IJB-C, respectively, while transforming the association between computational cost and the number of identities from linear to logarithmic. See code at https://github.com/msed-Ebrahimi/GIF

Figures

Figures reproduced from arXiv: 2505.03012 by the authors.

Figure 1
Figure 1. a) Illustrating the growth in the number of identities in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. a) We first convert scalar labels to identity codes. We use [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Left) Overview of proposed tokenization. We position each [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The maximum ↑, minimum ↑, and mean ↑ pairwise co￾sine distance among Softmax centroids of Fully Connected (FC) ArcFace (FC) [11], PFC [2], and the code vectors hi of our pro￾posal when d = 512. More separation among hi reflects better metric-space exploitation, leading…
Figure 5
Figure 5. Figure 5: (a, b) GPU memory consumption and Training speed comparison between FC, PFC and GIF: GIF significantly improves training [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: a) Showing the value of the λ for each token index in different scenarios. b) GIF performance is the best when the bal￾ancing factor of tokens, i.e., λ, is uniform across tokens [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: GPU memory consumption (a) and Training speed (b) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

81 extracted references · 57 canonical work pages

  1. [1]

    Partial fc: Training 10 million identities on a single ma- chine

    Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single ma- chine. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 1445–1449, 2021. 1, 2

  2. [2]

    Killing two birds with one stone: Efficient and robust training of face recognition cnns by partial fc

    Xiang An, Jiankang Deng, Jia Guo, Ziyong Feng, XuHan Zhu, Jing Yang, and Tongliang Liu. Killing two birds with one stone: Efficient and robust training of face recognition cnns by partial fc. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4042–4051, 2022. 1, 2, 3, 5, 6, 7, 8

  3. [3]

    Umdfaces: An annotated face dataset for training deep networks

    Ankan Bansal, Anirudh Nanduri, Carlos D Castillo, Rajeev Ranjan, and Rama Chellappa. Umdfaces: An annotated face dataset for training deep networks. In 2017 IEEE interna- tional joint conference on biometrics (IJCB), pages 464–473. IEEE, 2017. 2

  4. [4]

    Monotone funktionen, stieltjessche in- tegrale und harmonische analyse

    Salomon Bochner. Monotone funktionen, stieltjessche in- tegrale und harmonische analyse. Mathematische Annalen, 108(1):378–410, 1933. 4

  5. [5]

    Vggface2: A dataset for recognising faces across pose and age

    Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and An- drew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international con- ference on automatic face & gesture recognition (FG 2018), pages 67–74. IEEE, 2018. 2

  6. [6]

    Unsupervised learning of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Pi- otr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Ad- vances in neural information processing systems , 33:9912– 9924, 2020. 4

  7. [7]

    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 Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2

  8. [8]

    A generative approach for wikipedia-scale visual entity recognition

    Mathilde Caron, Ahmet Iscen, Alireza Fathi, and Cordelia Schmid. A generative approach for wikipedia-scale visual entity recognition. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 17313–17322, 2024. 2, 3, 8

Show all 81 references
  1. [9]

    Data uncertainty learning in face recognition

    Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710–5719, 2020. 7

  2. [10]

    Autoregressive entity retrieval

    Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. Autoregressive entity retrieval. arXiv preprint arXiv:2010.00904, 2020. 2, 4

  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 con- ference on computer vision and pattern recognition , pages 4690–4699, 2019. 1, 2, 3, 5, 6, 7, 8

  4. [12]

    Variational prototype learning for deep face recognition

    Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep face recognition. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 11906–11915, 2021. 2, 3, 1

  5. [13]

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

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6

  6. [14]

    Uniformface: Learn- ing deep equidistributed representation for face recognition

    Yueqi Duan, Jiwen Lu, and Jie Zhou. Uniformface: Learn- ing deep equidistributed representation for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3415–3424, 2019. 4

  7. [15]

    Explor- ing deep neural networks via layer-peeled model: Minority collapse in imbalanced training

    Cong Fang, Hangfeng He, Qi Long, and Weijie J Su. Explor- ing deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences, 118(43):e2103091118, 2021. 2, 1

  8. [16]

    Ms-celeb-1m: A dataset and benchmark for large-scale face recognition

    Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11-14, 2016, Proceedings, Part III 14, pages 8...

  9. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 6

  10. [18]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 9729–9738, 2020. 2

  11. [19]

    Fix your clas- sifier: the marginal value of training the last weight layer

    Elad Hoffer, Itay Hubara, and Daniel Soudry. Fix your clas- sifier: the marginal value of training the last weight layer. arXiv preprint arXiv:1801.04540, 2018. 3

  12. [20]

    Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities

    Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming- Wei Chang. Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities. In Proceedings of the IEEE/CVF International Conference on Computer...

  13. [21]

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

    Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database forstudying face recognition in unconstrained environments. In Workshop on faces in’Real-Life’Images: detection, align- ment, and recognition, 2008. 6

  14. [22]

    Curricularface: adaptive curriculum learning loss for deep face recognition

    Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 5...

  15. [23]

    Max- imum class separation as inductive bias in one matrix

    Tejaswi Kasarla, Gertjan Burghouts, Max van Spengler, Elise van der Pol, Rita Cucchiara, and Pascal Mettes. Max- imum class separation as inductive bias in one matrix. Advances in Neural Information Processing Systems , 35: 19553–19566, 2022. 4

  16. [24]

    Introduction to pytorch

    Nikhil Ketkar, Jojo Moolayil, Nikhil Ketkar, and Jojo Moolayil. Introduction to pytorch. Deep learning with python: learn best practices of deep learning models with PyTorch, pages 27–91, 2021. 2

  17. [25]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661–18673,

  18. [26]

    Adaface: Quality adaptive margin for face recognition

    Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18750–18759, 2022. 1, 2, 6, 7

  19. [27]

    Groupface: Learning latent groups and con- structing group-based representations for face recognition

    Yonghyun Kim, Wonpyo Park, Myung-Cheol Roh, and Jongju Shin. Groupface: Learning latent groups and con- structing group-based representations for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5621–5630, 2020. 7

  20. [28]

    Broad- face: Looking at tens of thousands of people at once for face recognition

    Yonghyun Kim, Wonpyo Park, and Jongju Shin. Broad- face: Looking at tens of thousands of people at once for face recognition. In European Conference on Computer Vision , pages 536–552. Springer, 2020. 7

  21. [29]

    Recurrent pixel embed- ding for instance grouping

    Shu Kong and Charless C Fowlkes. Recurrent pixel embed- ding for instance grouping. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 9018–9028, 2018. 4

  22. [30]

    Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text process- ing

    T Kudo. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text process- ing. arXiv preprint arXiv:1808.06226, 2018. 2, 4

  23. [31]

    Subword regularization: Improving neural net- work translation models with multiple subword candidates

    Taku Kudo. Subword regularization: Improving neural net- work translation models with multiple subword candidates. arXiv preprint arXiv:1804.10959, 2018. 4

  24. [32]

    Dynamic class queue for large scale face recognition in the wild

    Bi Li, Teng Xi, Gang Zhang, Haocheng Feng, Junyu Han, Jingtuo Liu, Errui Ding, and Wenyu Liu. Dynamic class queue for large scale face recognition in the wild. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3763–3772, 2021. 1, 2, 3, 6, 7

  25. [33]

    Virtual fully- connected layer: Training a large-scale face recognition dataset with limited computational resources

    Pengyu Li, Biao Wang, and Lei Zhang. Virtual fully- connected layer: Training a large-scale face recognition dataset with limited computational resources. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13315–13324, 2021. 1, 2, 3, 6

  26. [34]

    Con- trollable and guided face synthesis for unconstrained face recognition

    Feng Liu, Minchul Kim, Anil Jain, and Xiaoming Liu. Con- trollable and guided face synthesis for unconstrained face recognition. In European Conference on Computer Vision , pages 701–719. Springer, 2022. 7

  27. [35]

    Sphereface: Deep hypersphere embedding for face recognition

    Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Sphereface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212–220,

  28. [36]

    Cosine normalization: Using cosine similarity instead of dot product in neural networks

    Chunjie Luo, Jianfeng Zhan, Xiaohe Xue, Lei Wang, Rui Ren, and Qiang Yang. Cosine normalization: Using cosine similarity instead of dot product in neural networks. InArtifi- cial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neura...

  29. [37]

    Iarpa janus benchmark-c: Face dataset and protocol

    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 2018 Inter- national Conference on Biometrics (ICB) , pages 158–165. I...

  30. [38]

    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 con- ference on computer vision and pattern recognition , pages 14225–14234, 2021. 7

  31. [39]

    Distance-based image classification: Gen- eralizing to new classes at near-zero cost

    Thomas Mensink, Jakob Verbeek, Florent Perronnin, and Gabriela Csurka. Distance-based image classification: Gen- eralizing to new classes at near-zero cost. IEEE transactions on pattern analysis and machine intelligence , 35(11):2624– 2637, 2013. 3

  32. [40]

    Hyper- spherical prototype networks

    Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyper- spherical prototype networks. Advances in neural informa- tion processing systems, 32, 2019. 1, 3, 4, 6, 8

  33. [41]

    Distributed representations of words and phrases and their compositionality

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural in- formation processing systems, 26, 2013. 3

  34. [42]

    Agedb: the first manually collected, in-the-wild age database

    Stylianos Moschoglou, Athanasios Papaioannou, Chris- tos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops , pa...

  35. [43]

    No fuss distance met- ric learning using proxies

    Yair Movshovitz-Attias, Alexander Toshev, Thomas K Le- ung, Sergey Ioffe, and Saurabh Singh. No fuss distance met- ric learning using proxies. In Proceedings of the IEEE in- ternational conference on computer vision , pages 360–368,

  36. [44]

    Level play- ing field for million scale face recognition

    Aaron Nech and Ira Kemelmacher-Shlizerman. Level play- ing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7044–7053, 2017. 2

  37. [45]

    Deep face recognition

    Omkar Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In BMVC 2015-Proceedings of the British Machine Vision Conference 2015 . British Machine Vision Association, 2015. 2

  38. [46]

    Synface: Face recognition with syn- thetic data

    Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with syn- thetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880–10890, 2021. 1

  39. [47]

    Learn- ing 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. Learn- ing transferable visual models from natural language super- vision. 2021. 2, 3, 4, 8

  40. [48]

    Recom- mender systems with generative retrieval

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunan- dan H Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q Tran, Jonah Samost, et al. Recom- mender systems with generative retrieval. arXiv preprint arXiv:2305.05065, 2023. 2

  41. [49]

    Deep convolutional neu- ral networks for image classification: A comprehensive re- view

    Waseem Rawat and Zenghui Wang. Deep convolutional neu- ral networks for image classification: A comprehensive re- view. Neural computation, 29(9):2352–2449, 2017. 3

  42. [50]

    A quality aware sample-to-sample comparison for face recognition

    Mohammad Saeed Ebrahimi Saadabadi, Sahar Rahimi Malakshan, Ali Zafari, Moktari Mostofa, and Nasser M Nasrabadi. A quality aware sample-to-sample comparison for face recognition. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 6129–...

  43. [51]

    Distributing many points on a sphere

    Edward B Saff and Amo BJ Kuijlaars. Distributing many points on a sphere. The mathematical intelligencer , 19:5– 11, 1997. 4

  44. [52]

    Facenet: A unified embedding for face recognition and clus- tering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clus- tering. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 815–823, 2015. 1, 2, 6, 7

  45. [53]

    Frontal to profile face verification in the wild

    Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1–9. IEEE, 2016. 6

  46. [54]

    Neural machine translation of rare words with subword units

    Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 4

  47. [55]

    Equiangular basis vectors

    Yang Shen, Xuhao Sun, and Xiu-Shen Wei. Equiangular basis vectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11755– 11765, 2023. 1, 3

  48. [56]

    Multi- attention multi-class constraint for fine-grained image recog- nition

    Ming Sun, Yuchen Yuan, Feng Zhou, and Errui Ding. Multi- attention multi-class constraint for fine-grained image recog- nition. In Proceedings of the European Conference on Com- puter Vision (ECCV), pages 805–821, 2018. 2

  49. [57]

    On the origin of num- ber and arrangement of the places of exit on the surface of pollen-grains

    Pieter Merkus Lambertus Tammes. On the origin of num- ber and arrangement of the places of exit on the surface of pollen-grains. Recueil des travaux botaniques n ´eerlandais, 27(1):1–84, 1930. 4

  50. [58]

    Transformer memory as a differentiable search index

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. Transformer memory as a differentiable search index. Advances in Neural Information Processing Systems,

  51. [59]

    Normface: L2 hypersphere embedding for face veri- fication

    Feng Wang, Xiang Xiang, Jian Cheng, and Alan Loddon Yuille. Normface: L2 hypersphere embedding for face veri- fication. In Proceedings of the 25th ACM international con- ference on Multimedia, pages 1041–1049, 2017. 2

  52. [60]

    The devil of face recognition is in the noise

    Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, and Chen Change Loy. The devil of face recognition is in the noise. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 765–780,

  53. [61]

    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. InCVPR,

  54. [62]

    Git: A generative image-to-text transformer for vision and language

    Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 4

  55. [63]

    An efficient training approach for very large scale face recognition

    Kai Wang, Shuo Wang, Panpan Zhang, Zhipeng Zhou, Zheng Zhu, Xiaobo Wang, Xiaojiang Peng, Baigui Sun, Hao Li, and Yang You. An efficient training approach for very large scale face recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

  56. [64]

    Understanding contrastive representation learning through alignment and uniformity on the hypersphere

    Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929–9939. PMLR, 2020. 4

  57. [65]

    Visual recognition with deep nearest centroids

    Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids. arXiv preprint arXiv:2209.07383, 2022. 3

  58. [66]

    Sphereface2: Binary classification is all you need for deep face recognition

    Yandong Wen, Weiyang Liu, Adrian Weller, Bhiksha Raj, and Rita Singh. Sphereface2: Binary classification is all you need for deep face recognition. arXiv preprint arXiv:2108.01513, 2021. 1, 2

  59. [67]

    Iarpa janus benchmark-b face dataset

    Cameron Whitelam, Emma Taborsky, Austin Blanton, Bri- anna Maze, Jocelyn Adams, Tim Miller, Nathan Kalka, Anil K Jain, James A Duncan, Kristen Allen, et al. Iarpa janus benchmark-b face dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition w...

  60. [68]

    Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable clas- sifier at the end of deep neural network? Advances in neural information processing systems, 35:37991–38002, 20...

  61. [69]

    Neural collapse with normalized features: A ge- ometric analysis over the riemannian manifold

    Can Yaras, Peng Wang, Zhihui Zhu, Laura Balzano, and Qing Qu. Neural collapse with normalized features: A ge- ometric analysis over the riemannian manifold. Advances in neural information processing systems, 35:11547–11560,

  62. [70]

    Learn- ing face representation from scratch

    Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learn- ing face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 1, 2

  63. [71]

    Distribution alignment: A unified frame- work for long-tail visual recognition

    Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified frame- work for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361–2370, 2021. 2

  64. [72]

    Accelerated training for massive classification via dynamic class selection

    Xingcheng Zhang, Lei Yang, Junjie Yan, and Dahua Lin. Accelerated training for massive classification via dynamic class selection. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2

  65. [73]

    Adacos: Adaptively scaling cosine logits for effec- tively learning deep face representations

    Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hong- sheng Li. Adacos: Adaptively scaling cosine logits for effec- tively learning deep face representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10823–10832, 2019. 3

  66. [74]

    Cross-pose lfw: A database for studying cross-pose face recognition in un- constrained environments

    Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-pose face recognition in un- constrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5:7, 2018. 6

  67. [75]

    Cross-age lfw: A database for studying cross-age face recognition in un- constrained environments

    Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in un- constrained environments. arXiv preprint arXiv:1708.08197,

  68. [76]

    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, Dalong Du, et al. WebFace260M: A benchmark unveiling the power of million-scale deep face recognition. In CVPR,

  69. [78]

    CE Derivative Considering layer-peeled model to make a tractable analysis [15, 68], the gradient of Equation 1 w.r.t. the wj is: ∂LCE ∂wj = nX i=1 [−( 1−pj(zi))ziδ(j,yi) +pj(zi)zi( 1−δ(j,yi))], (11) herepj(z) is the predicted probability that z = Fθ(x) be- longs to the j-th cl...

  70. [79]

    Ablation on λ Here, we examine the impact of varying each λj on the training process. Each λj quantifies the relative impor- tance of the j-th token during training, where a higher λj indicates greater impact of the corresponding token, and a lowerλj suggests less importance. ...

  71. [80]

    As demonstrated in Figure 7a, even with the number of identities reaching 64 million, the GPU memory usage remains significantly lower than the OOM threshold

    Code Vector Optimization Cost Here, we investigate the GPU memory consumption associ- ated with the optimization of code vectors. As demonstrated in Figure 7a, even with the number of identities reaching 64 million, the GPU memory usage remains significantly lower than the OOM...

  72. [81]

    Replacing CLIP with DINO In this study, we explore the sensitivity of GIF to changes in the model used for initializing code vectors. We con- Method Train SetLFW CPLFW CALFW CFP-FP Age-DB IJB-B IJB-C GIF (DINO)MS1MV299.85 94.4796.75 98.75 98.6794.9896.80 GIF (CLIP)MS1MV299.859...

  73. [2021]

    1, 2, 3, 5 GIF: Generative Inspiration for Face Recognition at Scale Supplementary Material

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.