Generalized Face Discovery (GFD)

• Generalized Face Discovery (GFD) is an open-world framework that combines traditional face identification with novel identity discovery in unlabeled data.
• It employs dynamic prefix generation with a frozen Vision Transformer and landmark-guided patch extraction to tailor features per instance.
• FaceGCD, a leading method in GFD, demonstrates superior performance by balancing supervised and unsupervised contrastive learning for known and unknown faces.

Generalized Face Discovery (GFD) is an open-world face recognition task that unifies traditional face identification with generalized category discovery. It requires a model to recognize both labeled and unlabeled known identities while simultaneously discovering previously unseen identities in unlabeled data. In the formulation introduced with "FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation" (Oh et al., 30 Jul 2025), GFD is motivated by the need to recognize and differentiate among both familiar and unfamiliar faces, and it is presented as a setting in which the high cardinality and fine-grained nature of face identities render existing generalized category discovery methods ineffective.

1. Formal task definition

GFD is defined over a labeled training set and an unlabeled set. Let

$\mathcal{D}_L = \{(x_i^L, y_i^L)\}_{i=1}^{N_L}$

be the labeled training set, where $y_i^L \in \mathcal{C}_L$ is one of the known face IDs, and let

$\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$

be the unlabeled set, where each $x_j^U$ either belongs to a known ID in $\mathcal{C}_L$ or to an unknown ID in $\mathcal{C}_N$. The identity sets are

$\mathcal{C}_L = \{c_1,\ldots,c_{|\mathcal{C}_L|}\}$

for labeled known IDs, and

$$\mathcal{C}_N = \{c_{|\mathcal{C}_L|+1},\ldots,c_{|\mathcal{C}_L|+|\mathcal{C}_N|}\}$$

for unknown IDs (Oh et al., 30 Jul 2025).

The learning objective is to obtain a mapping $f(\cdot)$ such that, for any test image $x$, the model either classifies $y_i^L \in \mathcal{C}_L$0 into one of the $y_i^L \in \mathcal{C}_L$1 known IDs if $y_i^L \in \mathcal{C}_L$2, or assigns $y_i^L \in \mathcal{C}_L$3 to a newly discovered cluster in $y_i^L \in \mathcal{C}_L$4 when $y_i^L \in \mathcal{C}_L$5. In this sense, GFD explicitly combines closed-set face identification on $y_i^L \in \mathcal{C}_L$6 with cluster discovery on unknown identities $y_i^L \in \mathcal{C}_L$7 (Oh et al., 30 Jul 2025).

2. Position within face analysis and open-world recognition

GFD occupies an intermediate position between closed-set face identification and fully unsupervised face clustering. Closed-set identification assumes that all test identities belong to a predefined label space, whereas clustering methods operate without that assumption and organize images by affinity structure. GFD requires both capabilities in a single model: maintaining discrimination over known IDs while discovering unknown ones.

This distinction is important because large-scale face clustering methods such as Ada-NETS are designed for entirely unlabeled settings. Ada-NETS builds a graph from pairwise feature affinities, makes no strong closed-set assumption, and does not require a pre-known number of clusters; during inference it thresholds GCN similarities and runs union-find to form connected components, so unseen identities are naturally placed in their own components if they lack high-affinity neighbours (Wang et al., 2022). GFD differs by requiring explicit recognition of known IDs in addition to discovery of new IDs.

The face domain also appears to resist direct transfer of generic generalized category discovery methods. The GFD formulation states that the high cardinality and fine-grained nature of face IDs render existing GCD approaches ineffective (Oh et al., 30 Jul 2025). A related result appears in facial expression analysis: "DIG-FACE: De-biased Learning for Generalized Facial Expression Category Discovery" reports that generalized category discovery methods for natural images show compromised performance on generalized facial expression category discovery, and attributes this to implicit and explicit biases (Luo et al., 2024). This suggests that face-related discovery tasks are sufficiently specialized to require domain-specific modeling.

3. FaceGCD and dynamic prefix generation

The first method proposed specifically for GFD is FaceGCD, which uses dynamic prefix generation to construct instance-specific feature extractors. Its backbone is a frozen Vision Transformer, specifically ViT-B/16 pretrained via DINO on MS1MV3. Given an input face $y_i^L \in \mathcal{C}_L$8, the model first extracts $y_i^L \in \mathcal{C}_L$9 landmark-guided patches and obtains layer-wise “identity-level” features $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$0 for $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$1 (Oh et al., 30 Jul 2025).

At each Transformer layer $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$2, FaceGCD injects two small prefix sets, $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$3 and $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$4, into the key and value inputs of self-attention: $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$5 while $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$6 remains unchanged. The prefixes are not static prompts. Instead, they are generated on the fly by a lightweight 2-layer MLP hypernetwork $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$7, which produces per-layer generator parameters

$\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$8

These parameters define bottleneck projections $\mathcal{D}_U = \{x_j^U\}_{j=1}^{N_U}$9. With randomly initialized prefix tokens $x_j^U$0, the generated prefixes are

$x_j^U$1

$x_j^U$2

where $x_j^U$3 is prefix length, $x_j^U$4 the number of heads, and $x_j^U$5 the per-head dimension (Oh et al., 30 Jul 2025).

The central design claim is that conditioning the key and value prefix generators on per-instance features $x_j^U$6 tailors the feature extractor to each image. In the discussion accompanying the method, this is described as instance-specific adaptation that captures subtle identity cues among visually similar faces without relying on high-capacity static models.

4. Optimization objectives and evaluation protocol

FaceGCD trains only the hypernetwork $x_j^U$7, the final DINO head $x_j^U$8, and the last layer weights of the ViT. For a batch $x_j^U$9, the total semi-supervised contrastive loss is

$\mathcal{C}_L$0

where $\mathcal{C}_L$1 balances supervised and unsupervised terms, and $\mathcal{C}_L$2 weights the supervised loss (Oh et al., 30 Jul 2025).

The unsupervised contrastive term is

$\mathcal{C}_L$3

which encourages each sample to match its augmentation and repel other instances in the batch. The supervised contrastive term is

$\mathcal{C}_L$4

where $\mathcal{C}_L$5 contains other samples sharing the same known ID label; this pulls together same-ID features and pushes apart different IDs (Oh et al., 30 Jul 2025).

The benchmarks are constructed from YouTube Faces and CASIA-WebFaces, with six GFD settings using $\mathcal{C}_L$6 total IDs. For each split, $\mathcal{C}_L$7 of IDs are known, with half of their training images labeled, and $\mathcal{C}_L$8 are unknown with no labels. On YTF 1000, the benchmark contains 500 known IDs, 500 unknown IDs, approximately $\mathcal{C}_L$9K training images, and approximately $\mathcal{C}_N$0K test images. Evaluation uses Clustering Accuracy (ACC), with Hungarian assignment reported separately on Known, Novel, and All, and Nearest-Neighbor Consistency (NNC), which measures the average fraction of $\mathcal{C}_N$1-nearest neighbours sharing the same pseudo-label or label, indicating feature separability (Oh et al., 30 Jul 2025).

5. Empirical results and ablations

Across all six GFD settings, FaceGCD is reported to outperform existing GCD baselines—GCD, SimGCD, PromptCAL, and CMS—as well as a strong ArcFace baseline, on the GFD task (Oh et al., 30 Jul 2025). On YTF 1000, the All ACC values are 82.7% for FaceGCD, 75.0% for CMS, 75.0% for PromptCAL, 72.8% for SimGCD, and 70.4% for GCD. For Known ACC, the corresponding values are 93.2% for FaceGCD, 86.1% for PromptCAL, and 95.1% for SimGCD. For Novel ACC, FaceGCD achieves 72.1%, compared with 63.8% for PromptCAL and 50.1% for SimGCD. On the same benchmark, NNC is 91.1% for FaceGCD, versus 82.7% for ArcFace and 83.6% for GCD (Oh et al., 30 Jul 2025).

These numbers are notable because they show that GFD performance cannot be reduced to known-ID accuracy alone. On YTF 1000, SimGCD attains a higher Known ACC than FaceGCD, but FaceGCD achieves the highest All ACC and a substantially higher Novel ACC. This indicates that the discovery component materially affects aggregate performance in the open-world setting.

The ablation studies isolate the role of dynamic prefixes. On YTF 1000, reducing prefix length to 5 yields All 80.9%, Known 91.1%, and Novel 70.4%, while prefix length 20 yields All 82.7%. Replacing the hypernetwork with a static prefix generator yields All 76.9%, a drop of 6.0 points, despite using 30× more parameters. A static prefix pool yields All 42.6%, a drop of 40.3 points. The discussion attributes this to the inability of static prompts to capture instance-level variation in high-cardinality settings. The added hypernetwork and prefix modules contribute only approximately 6.6% extra parameters, or about 13.8M, in contrast to hundreds of millions in static prompt MLPs (Oh et al., 30 Jul 2025).

6. Relation to adjacent discovery paradigms

GFD is closely related to, but distinct from, two neighboring paradigms: large-scale face clustering and generalized facial expression discovery.

In face clustering, Ada-NETS addresses web-scale unlabeled data by transforming features into a structure space whose similarity kernel combines reciprocal-neighbour Jaccard similarity with cosine similarity,

$\mathcal{C}_N$2

using $\mathcal{C}_N$3, and then applying adaptive neighbour discovery before a two-layer GCN. On MS-Celeb-1M, this design improves the face-graph signal-to-noise ratio from 1.62 without structure space or adaptive discovery to 21.67 when both are enabled, while BCubed $\mathcal{C}_N$4 rises from 77.25 to 91.40. Ada-NETS is explicitly presented as well-suited to massive, entirely unlabeled, in-the-wild web imagery (Wang et al., 2022). The misconception that GFD is simply large-scale clustering is therefore inaccurate: clustering addresses discovery without labeled known-ID recognition, whereas GFD requires both.

In facial expression analysis, DIG-FACE defines Generalized Facial Expression Category Discovery, where a single model must continue to classify known expressions and discover new ones. It attributes failures of generic generalized category discovery methods to implicit bias, arising from a distributional gap between old and new categories, and explicit bias, arising from shifted preference on facial change characteristics, and addresses these using adversarial discrepancy minimization together with sample-level, triplet-level, and distribution-level refinement (Luo et al., 2024). Because DIG-FACE concerns expression categories rather than identity, it is not a GFD method. A plausible implication is that open-world discovery problems on faces may repeatedly require domain-specific mechanisms beyond generic image-category discovery.

7. Limitations, open questions, and research directions

The stated limitations of FaceGCD are architectural and operational. The method relies on a pretrained landmark detector and a pretrained ViT, and its performance may depend on landmark quality (Oh et al., 30 Jul 2025). This dependence is particularly relevant because the pipeline begins from landmark-guided patch extraction; if landmarks are unstable, the identity-level features conditioning the hypernetwork may also be affected.

The open questions identified for GFD concern deployment beyond the static benchmark setting. Real-world streams may require continual or incremental clustering of new IDs. Future work is suggested in four directions: end-to-end learning of patch selection, dynamic expansion of the unknown-ID set over time, domain adaptation to novel capture conditions, and integration with open-set recognition thresholds (Oh et al., 30 Jul 2025).

Taken together, these limitations indicate that GFD should be understood not as a solved recognition problem but as a formalization of open-world identity understanding in which closed-set classification and unsupervised discovery are jointly optimized. The present formulation establishes the task, demonstrates that existing generalized category discovery methods are insufficient in the fine-grained identity regime, and provides a concrete dynamic-prefix baseline for subsequent work (Oh et al., 30 Jul 2025).