VFace10K: Synthetic Face Dataset for FR

• VFace10K is a synthetic face recognition dataset comprising 10,000 identities with 50 images per identity, emerging from both Vec2Face+ and VariFace methodologies.
• The VariFace pipeline employs a two-stage conditional diffusion process to ensure inter-class separability, controlled intra-class diversity, and demographic fairness.
• Vec2Face+ uses feature-to-image generation with techniques like AttrOP and LoRA-based pose control to enhance identity consistency and overall synthetic image quality.

VFace10K denotes a 10,000-identity synthetic face-recognition training set in two closely related senses in recent literature. In "Vec2Face+ for Face Dataset Generation," it is the explicit name of a dataset generated by Vec2Face+, with 50 images per identity at $112\times112\times3$ resolution, designed to combine inter-class separability, intra-class attribute variation, and intra-class identity consistency (Wu et al., 23 Jul 2025). In "VariFace: Fair and Diverse Synthetic Dataset Generation for Face Recognition," the name itself is not introduced; however, the constrained VariFace configuration of $10\text{K}$ identities $\times$ 50 images per identity is stated to map directly to what may be meant by “VFace10K,” again totaling $0.5\text{M}$ synthetic images (Yeung et al., 2024). The term therefore has an ambiguous but technically traceable usage: one explicit and one retrospective.

1. Nomenclature, scope, and motivation

The literature represented here uses VFace10K in two distinct ways. One is formal: VFace10K is a named dataset created with Vec2Face+. The other is descriptive: “VFace10K” refers to the constrained-size VariFace setting with $10{,}000$ synthetic identities and 50 images per identity. In both cases, the target application is face recognition (FR) training on synthetic data rather than on large-scale web-scraped biometric corpora (Yeung et al., 2024, Wu et al., 23 Jul 2025).

Usage in literature	Meaning	Composition
VFace10K in Vec2Face+	Explicit dataset name	10,000 identities; 50 images per identity; $\sim 0.5\text{M}$ images
“VFace10K” as applied to VariFace	Constrained VariFace configuration	10,000 identities; 50 images per identity; $0.5\text{M}$ images

The shared motivation is the same broad problem setting. Real, web-scraped face datasets raise privacy and legal concerns, including GDPR Article 9 and EU AI Act Article $5(1)(e)$, and they often exhibit demographic imbalance, bias, and limited pose or illumination variability. Synthetic generation is presented as a route to scalable and controllable FR training data. The two lines of work differ in emphasis. VariFace centers fairness, demographic balancing, and controlled diversity. Vec2Face+ centers the claim that prior synthetic FR datasets neglected intra-class identity consistency even when pursuing large inter-class separability and large intra-class variation.

2. VariFace 10K$\times$50: diffusion-based fair and diverse synthesis

In the VariFace framework, the $10\text{K}\times 50$ configuration is produced by a two-stage conditional diffusion pipeline built on HDiT, with automatic filtering and explicit guidance or conditioning for interclass and intraclass diversity (Yeung et al., 2024). Stage 1 generates demographically balanced synthetic identities. Its inputs are initial demographic labels from CLIP—race, gender, and age—and identity embeddings $10\text{K}$0 from a pretrained FR model, IResNet-100. The initial labels are refined by Face Recognition Consistency (FRC), and sampling is conditioned on race and gender. Face Vendi Score Guidance (FVSG) is then applied during diffusion sampling to improve interclass diversity in FR embedding space. Outputs inconsistent with the intended demographic label are removed via CLIP-FRC, and quality filtering is performed with CLIB-FIQA at threshold $10\text{K}$1.

The race classes are Caucasian, Asian, Indian, and African, aligned to RFW categories, and gender is binary Male/Female. Stage 2 then generates multiple images per identity with diverse pose and appearance while preserving identity. It takes as input normalized mean identity embeddings from Stage 1, age labels from CLIP, and divergence scores (DS) computed from FR embeddings. The conditioning variables in Stage 2 are identity embedding, age, and DS. Age is sampled uniformly, $10\text{K}$2, and DS is sampled uniformly as $10\text{K}$3 by default. Identity preservation is enforced post hoc with a cosine-similarity threshold of $10\text{K}$4 between the Stage 1 base embedding and Stage 2 images.

FRC is a label-refinement rule rather than a learned loss. It uses cosine similarity in FR embedding space,

$10\text{K}$5

selects top-$10\text{K}$6 neighbors for each image, and reassigns the image’s race or gender label to the majority label among those neighbors. FVSG uses the Vendi Score at batch level,

$10\text{K}$7

with guidance loss

$10\text{K}$8

The diffusion update uses

$10\text{K}$9

$\times$0

$\times$1

Empirically, higher guidance scale increases diversity but can increase artifacts: guidance scale $\times$2 gives average cosine similarity $\times$3 and high-quality fraction $\times$4; scale $\times$5 gives $\times$6 and $\times$7; scale $\times$8 gives $\times$9 and $0.5\text{M}$0.

Intraclass diversity is controlled through divergence-score conditioning. For identity $0.5\text{M}$1, the prototypical embedding is

$0.5\text{M}$2

and the divergence score of sample $0.5\text{M}$3 is

$0.5\text{M}$4

Lower DS encourages larger variation, whereas excessively low DS risks identity loss. The reported best FR performance occurs for $0.5\text{M}$5; $0.5\text{M}$6 and $0.5\text{M}$7 both degrade performance.

A notable property of this pipeline is that the constrained $0.5\text{M}$8 setting is only one operating point. The same framework is reported in unconstrained settings up to $0.5\text{M}$9 identities with 100 images per identity. This suggests that “VFace10K” in the VariFace sense is best understood as a small-scale instance of a scalability study rather than as a separately packaged dataset.

3. VFace10K in Vec2Face+: feature-to-image generation with identity-consistency control

In the Vec2Face+ formulation, VFace10K is generated by a feature-to-image model trained on WebFace4M, comprising 4M images, without identity labels (Wu et al., 23 Jul 2025). The model maps pretrained FR features directly to images. Identity vectors are sampled in a 512-dimensional embedding space defined by an ArcFace-R100 backbone, using cosine similarity on the hypersphere as the relevant metric. The sampling heuristic restricts cosine similarity between distinct sampled identity vectors to be small, empirically $10{,}000$0, and the paper notes that in high dimensions one can “easily sample 4M identity vectors” with pairwise cosine $10{,}000$1.

Generation proceeds through three bases. The first is random feature perturbation: per-identity vectors are perturbed and decoded into images. The second is AttrOP, a gradient-based attribute operation that increases general attribute variation, including pose-driven variation. The third is LoRA-based pose control conditioned on face landmarks. Final dataset assembly begins from a random base of 50 images per identity and replaces 40 of those images with AttrOP and LoRA pose-control outputs, yielding 50 images per identity with broad attribute and pose coverage and improved identity consistency.

The main Vec2Face+ architecture expands an input FR feature from shape $10{,}000$2 to $10{,}000$3 via two linear layers so that the representation aligns with ViT-B token shapes. An fMAE module randomly masks rows from the expanded feature map according to a trunc-normal mask ratio $10{,}000$4, encodes the partially masked representation, then fills masked rows with a condition before decoding. The image decoder uses 4 deconvolution layers to output $10{,}000$5 images. Vec2Face+ removes the patch-based GAN discriminator used by Vec2Face, reducing training time by $10{,}000$6 with no reported image-quality drop.

The loss suite is

$10{,}000$7

$10{,}000$8

$10{,}000$9

$\sim 0.5\text{M}$0

Identity consistency is enforced primarily at generation time rather than through an additional explicit training loss. Identity images are accepted only if $\sim 0.5\text{M}$1 and $\sim 0.5\text{M}$2, where $\sim 0.5\text{M}$3 is MagFace quality magnitude. Per-identity images generated from perturbed features are accepted only if $\sim 0.5\text{M}$4 and $\sim 0.5\text{M}$5. Perturbations use $\sim 0.5\text{M}$6 with $\sim 0.5\text{M}$7, and the embedding norm is later controlled to $\sim 0.5\text{M}$8.

AttrOP optimizes a latent feature vector toward target pose and quality: $\sim 0.5\text{M}$9 with $0.5\text{M}$0 as the generator, $0.5\text{M}$1 as SixDRepNet, $0.5\text{M}$2 as the MagFace magnitude estimator, and $0.5\text{M}$3 as ArcFace-R100. Pose targets use yaw $0.5\text{M}$4, and the optimization is iterated $0.5\text{M}$5 times. Because large target poses can be time-consuming and can induce identity drift, the paper introduces LoRA-based pose control. The LoRA parameterization is

$0.5\text{M}$6

with $0.5\text{M}$7, $0.5\text{M}$8, and $0.5\text{M}$9. The main model weights are frozen, low-rank adapters are inserted into selected layers, and a lightweight 4-layer CNN encodes a 5-landmark pose image. For 200K images, LoRA pose control requires $5(1)(e)$0 minutes on an NVIDIA L40S, versus $5(1)(e)$1 hours with AttrOP.

Two analytical quantities organize the paper’s evaluation of synthetic identity quality. Inter-class separability is

$5(1)(e)$2

where an identity is “well-separated” if its mean embedding has cosine similarity below $5(1)(e)$3 with every other identity. Identity consistency is $5(1)(e)$4, the average intra-class cosine similarity across image pairs within each identity, computed using ArcFace-R100. Correlation analyses reported in the paper indicate that separability helps until saturation near $5(1)(e)$5, whereas higher identity consistency more strongly correlates with better test accuracy.

4. Reported performance and scaling behavior

The two usages of VFace10K participate in different evaluation protocols and therefore should not be numerically conflated. VariFace reports on the Standard Benchmark—LFW, CFP-FP, CPLFW, AgeDB, and CALFW—and on RFW, with Real Gap defined as $5(1)(e)$6, where the real baseline is a model trained on CASIA-WebFace (Yeung et al., 2024). Vec2Face+ reports accuracy percentages on five standard sets—LFW, CFP-FP, CPLFW, AgeDB-30, and CALFW—plus Hadrian, Eclipse, IJBB, and IJBC (Wu et al., 23 Jul 2025).

System or usage	Scale	Headline result
VariFace constrained $5(1)(e)$7	$5(1)(e)$8 images	Standard Benchmark average $5(1)(e)$9; RFW average $\times$0
VariFace $\times$1	$\times$2 images	Standard Benchmark average $\times$3; RFW average $\times$4
Vec2Face+ VFace10K	$\times$5 images	Five-set average $\times$6
Vec2Face+ VFace100K	$\times$7 images	Five-set average $\times$8, above CASIA-WebFace $\times$9
Vec2Face+ VFace300K	$10\text{K}\times 50$0 images	Five-set average $10\text{K}\times 50$1, above CASIA-WebFace $10\text{K}\times 50$2

For constrained VariFace $10\text{K}\times 50$3, the CASIA-WebFace baseline on the Standard Benchmark is LFW $10\text{K}\times 50$4, CFP-FP $10\text{K}\times 50$5, CPLFW $10\text{K}\times 50$6, AgeDB $10\text{K}\times 50$7, CALFW $10\text{K}\times 50$8, average $10\text{K}\times 50$9. The corresponding VariFace scores are LFW $10\text{K}$00, CFP-FP $10\text{K}$01, CPLFW $10\text{K}$02, AgeDB $10\text{K}$03, CALFW $10\text{K}$04, average $10\text{K}$05, giving Real Gap $10\text{K}$06. On RFW, CASIA-WebFace achieves African $10\text{K}$07, Asian $10\text{K}$08, Caucasian $10\text{K}$09, Indian $10\text{K}$10, average $10\text{K}$11; VariFace $10\text{K}$12 achieves African $10\text{K}$13, Asian $10\text{K}$14, Caucasian $10\text{K}$15, Indian $10\text{K}$16, average $10\text{K}$17, giving Real Gap $10\text{K}$18. Relative to prior synthetic baselines in the constrained regime, the previous best Standard Benchmark average is $10\text{K}$19 from Vec2Face, whereas VariFace reports $10\text{K}$20, and on constrained RFW the previous best average is $10\text{K}$21 from Vec2Face versus $10\text{K}$22 for VariFace.

VariFace also reports unconstrained scaling. Standard Benchmark averages are $10\text{K}$23 for $10\text{K}$24 ($10\text{K}$25), $10\text{K}$26 for $10\text{K}$27 ($10\text{K}$28), $10\text{K}$29 for $10\text{K}$30 ($10\text{K}$31), and $10\text{K}$32 for $10\text{K}$33 ($10\text{K}$34). The corresponding Real Gaps are $10\text{K}$35, $10\text{K}$36, $10\text{K}$37, and $10\text{K}$38. On RFW, the same scales obtain averages $10\text{K}$39, $10\text{K}$40, $10\text{K}$41, and $10\text{K}$42, with Real Gaps $10\text{K}$43, $10\text{K}$44, $10\text{K}$45, and $10\text{K}$46. The paper states that from $10\text{K}$47 images onward, VariFace surpasses CASIA-WebFace across the Standard Benchmark, and at $10\text{K}$48 it sets a new state-of-the-art average of $10\text{K}$49.

For Vec2Face+ VFace10K, the five standard test-set accuracies are LFW $10\text{K}$50, CFP-FP $10\text{K}$51, CPLFW $10\text{K}$52, AgeDB-30 $10\text{K}$53, and CALFW $10\text{K}$54, with average $10\text{K}$55. The paper states that this is $10\text{K}$56 over the second-best synthetic dataset of similar size, $10\text{K}$57, while CASIA-WebFace has average $10\text{K}$58. Additional test sets give Hadrian $10\text{K}$59, Eclipse $10\text{K}$60, IJBB $10\text{K}$61 at $10\text{K}$62, and IJBC $10\text{K}$63 at $10\text{K}$64; CASIA-WebFace gives $10\text{K}$65, $10\text{K}$66, $10\text{K}$67, and $10\text{K}$68, respectively. Scaling yields VFace20K at $10\text{K}$69 images with average $10\text{K}$70, VFace100K at $10\text{K}$71 with $10\text{K}$72, and VFace300K at $10\text{K}$73 with $10\text{K}$74, exceeding CASIA-WebFace’s $10\text{K}$75. The reported gains diminish from 100K to 300K identities.

5. Fairness, bias, and edge cases

Fairness is an explicit design axis in VariFace. The generation process balances race and gender in Stage 1, uses RFW-aligned race classes, and evaluates subgroup face-verification accuracy on African, Asian, Caucasian, and Indian subsets (Yeung et al., 2024). In the constrained $10\text{K}$76 setting, VariFace exceeds the CASIA-WebFace baseline on African ($10\text{K}$77 vs. $10\text{K}$78), Asian ($10\text{K}$79 vs. $10\text{K}$80), and Indian ($10\text{K}$81 vs. $10\text{K}$82) while remaining slightly below on Caucasian ($10\text{K}$83 vs. $10\text{K}$84). Ablation results state that removing race or gender conditioning reduces RFW performance, notably for Asian, Indian, and African subgroups. Qualitatively, t-SNE plots show balanced clusters across race and gender for VariFace, including Indian clusters absent in CASIA-WebFace, whereas other synthetic methods often inherit CASIA-WebFace biases, such as dominant Caucasian clusters.

Vec2Face+ reports a different conclusion regarding bias. On BA-test and BFW, using $10\text{K}$85 at $10\text{K}$86 by demographic subgroup, models trained on synthetic identities show larger demographic disparities than models trained on real identities (Wu et al., 23 Jul 2025). The paper notes consistent patterns such as larger gender gaps within certain races, exemplified by BM versus BF, and observes that the lowest accuracies frequently occur in Asian subgroups while the highest often occur in Indian subgroups. Proposed mitigations—more balanced synthetic generation across demographic attributes, attribute-aware sampling or targets in AttrOP and LoRA conditioning, and bias-aware training objectives—are presented only as future directions.

An important misconception addressed by the Vec2Face+ work is that inter-class separability and broad attribute variation alone are sufficient for high-quality synthetic FR training data. The paper argues that high intra-class identity consistency is the missing factor in prior synthetic datasets, and its correlation study reports that separability helps until saturation near $10\text{K}$87, whereas higher identity consistency more strongly correlates with better test accuracy. The VariFace results point in a compatible direction: DS extremes hurt either identity preservation or diversity, and the best performance arises from a constrained diversity band $10\text{K}$88.

A second edge case is identical twins verification. On Twins-IND, where a trivial baseline is random guessing at $10\text{K}$89, only 1 of 11 synthetic training sets exceeds $10\text{K}$90; VFace300K reaches $10\text{K}$91, while VFace10K and most others are approximately $10\text{K}$92 (Wu et al., 23 Jul 2025). The stated implication is that synthetic identity formation based on large inter-class angular separation does not model real-world near-duplicate relationships such as twins. This suggests that synthetic identity spaces optimized for FR benchmarks can remain underconstrained with respect to biologically or socially similar identities.

6. Availability, training protocols, and practical status

Neither line of work presents VFace10K as a publicly specified benchmark package with the usual metadata. For VariFace, the paper does not provide a public download link, file format specification, explicit licensing or usage terms, or train/validation/test splits, and access instructions are not included (Yeung et al., 2024). For Vec2Face+ VFace10K, file formats are not specified, train/validation splits are not specified, all reported experiments train on the full dataset, and no download link or license is provided (Wu et al., 23 Jul 2025). In both cases, the literature demonstrates generation and use rather than release infrastructure.

The reported FR training protocols also differ. For FR models consuming VariFace, the evaluation backbone is IResNet-50, while IResNet-100 is used inside the generation pipeline and for FRC. The FR loss is ArcFace with scale $10\text{K}$93 and margin $10\text{K}$94. Optimization uses SGD with initial learning rate $10\text{K}$95, StepLR $10\text{K}$96 at epochs 24, 30, and 36, for 40 epochs, with batch size 256. Augmentations include horizontal flip, sharpness, autocontrast, equalize, normalization, and random erasing. The implementation follows InsightFace defaults except for scheduler and extra augmentations. For FR models trained on VFace datasets in Vec2Face+, the architecture is SE-IResNet50 with ArcFace loss, preprocessing resizes to $10\text{K}$97, augmentations include horizontal flip, random crop, low resolution, random erase, and photometric augmentations, and optimization uses SGD for 40 epochs with initial learning rate $10\text{K}$98 decayed at epochs 18, 28, and 35.

Identity leakage is addressed in both literatures, but with different protocols. VariFace reports a memorization test in which the maximum cosine similarity between $10\text{K}$99 VariFace identities and CASIA-WebFace identities is mostly around $\times$00, with very few cases above $\times$01, suggesting distinct identities. Vec2Face+ enforces a stronger exclusion rule during generation by dropping any generated image whose embedding has cosine similarity $\times$02 with any WebFace4M identity embedding. Both sets of results are intended to support the claim that the generated identities are not simple replicas of source identities, although both pipelines still depend on real datasets during model construction.

In practical use, the papers present VFace10K-class datasets as training resources for FR models trained purely on synthetic data, as pretraining corpora, as augmentation sources, and, in the VariFace case, as fairness-aware dataset constructions with controlled demographics. The principal limitations remain explicit. VariFace $\times$03 does not surpass CASIA-WebFace in the constrained setting even though it approaches parity, and its pipeline still uses CASIA-WebFace as the real source for training diffusion and FR components. Vec2Face+ reports stronger-than-real average accuracy only after scaling to VFace100K or VFace300K, while also documenting increased demographic bias and near-random performance on twins. Taken together, these results place VFace10K at an intermediate point in the synthetic-face-recognition literature: technically mature enough to support competitive FR training, but not yet a fully resolved substitute for real-identity data across fairness, release practice, and edge-case generalization.