The Surprising Geometry of Facial Identity: Are Face Embeddings Compatible Across Deep Neural Network Models?

Modern face recognition systems rely on a diverse ecosystem of deep neural network architectures. On one hand, we have domain-specific models such as ArcFace, AdaFace, and MagFace, meticulously trained on curated facial datasets with metric learning objectives. On the other hand, foundation models, pretrained on broad vision or vision-language tasks, increasingly demonstrate competence in biometric identification without ever being explicitly trained to distinguish faces. This architectural diversity raises a fundamental question about the nature of representation learning. Do these independently trained systems encode facial identity in fundamentally incompatible ways, or do they discover similar geometric structures in the high-dimensional space of human faces?

In their recent study, "Are Face Embeddings Compatible Across Deep Neural Network Models?", Rubab, Tong, and Ross from Michigan State University investigate this question through a rigorous geometric analysis of embedding spaces. Their findings challenge prior assumptions about the incompatibility of face embeddings and reveal surprising structural commonalities across model families.

The Mechanics of Cross-Model Alignment

The researchers approach this problem by treating face embeddings as point clouds in high-dimensional vector spaces. Given two distinct models, M1 and M2, they ask whether a simple transformation can align the embedding produced by M1 for a given face with the corresponding embedding from M2. Specifically, they investigate whether low-capacity affine mappings suffice to bridge the representational gap between models.

The technical methodology employs three distinct alignment strategies of increasing complexity. First, Procrustes alignment enforces orthogonal transformations, preserving the geometric structure while allowing rotation and scaling. Second, unconstrained linear regression provides greater flexibility through general linear transformations. Third, Ridge regression introduces regularization to prevent overfitting when learning the mapping. These methods operate under the assumption that if two models capture the same underlying identity manifold, they should differ primarily in coordinate parameterization rather than intrinsic geometry.

The evaluation protocol spans both face identification (1:N matching) and verification (1:1 comparison) tasks. The results demonstrate that these simple linear transformations substantially improve cross-model recognition accuracy over unaligned baselines. For instance, when querying with embeddings from BLIP-2 against a gallery of LLaVA embeddings, the unaligned baseline yields incorrect nearest neighbors, whereas linear alignment recovers the correct identity matches. This suggests that independently trained models, despite differing architectures, training objectives, and data regimes, encode facial identity in geometrically compatible ways.

Representational Convergence Across Architectures

Perhaps more intriguing than the alignment itself is the observation that these patterns generalize across datasets and reveal systematic structural properties. The researchers find that alignment transformations learned on one dataset transfer effectively to others, indicating that the geometric relationships between embedding spaces are not dataset-specific artifacts but reflect deeper representational properties.

The study reveals a hierarchy of compatibility across model families. Domain-specific models, trained explicitly to separate identities through metric learning, and foundation models, which encode visual structure through self-supervision or vision-language alignment, both encourage structured organization of the identity manifold. The authors interpret this as representational convergence: different learning objectives and architectural inductive biases nonetheless converge on similar geometric solutions for encoding facial identity.

This convergence has significant implications for our understanding of neural network learning dynamics. The fact that simple linear mappings suffice suggests that these models are parameterizing local tangent spaces of a common underlying manifold. From a manifold learning perspective, each model provides a different coordinate chart of the same facial identity space, and linear transformations approximate the Jacobian of the transition map between these charts. The observed directional asymmetry in target spaces, where mapping from model A to model B may differ in fidelity from the reverse mapping, further supports this differential geometric interpretation.

Beyond Alignment: Security, Interoperability, and Limitations

These findings carry profound implications for biometric system design and security that extend beyond the paper's immediate scope. Prior work by Bhatta et al. suggested that embeddings from independently trained instances of the same architecture are mutually incompatible, proposing this as a mechanism for implicit template revocability. The idea was that if embeddings are inherently incompatible, compromised templates could be revoked by simply switching to a different model instance without security concerns about cross-model matching.

However, if low-capacity linear transformations can align these spaces with high accuracy, the security properties of such revocability schemes require reevaluation. An adversary with knowledge of the alignment transformation could potentially map revoked templates into a new model's embedding space, undermining the protection mechanism. This suggests that template security strategies must account for the geometric compatibility demonstrated in this work, possibly requiring non-linear transformations or cryptographic protection rather than relying on assumed representational incompatibility.

From an interoperability standpoint, the ability to align embeddings across models enables novel ensemble architectures. Systems could combine the complementary strengths of domain-specific models, which excel at fine-grained discrimination, with foundation models, which generalize better to unconstrained conditions, without requiring costly retraining or joint optimization. The linear nature of the alignment makes this computationally efficient, requiring only a simple matrix multiplication at inference time.

Nevertheless, the study's focus on linear transformations presents both a strength and a limitation. While the efficacy of linear alignment demonstrates strong geometric similarity, it does not preclude the existence of residual non-linear structure that linear maps fail to capture. Future work should investigate whether non-linear alignment methods, such as those based on optimal transport or neural network-based transforms, could achieve even higher cross-model accuracy, or whether fundamental information loss in certain model pairs imposes hard limits on compatibility.

The Road Ahead

The convergence of facial representations across disparate neural architectures suggests that the geometry of human identity imposes strong constraints on learnable representations. This observation opens several avenues for future inquiry. Do similar alignment properties hold for other biometric modalities, such as iris or fingerprint recognition, where the underlying physical structures differ significantly from the continuous deformations of facial geometry? Can we characterize the specific inductive biases, whether architectural or objective-based, that promote or hinder representational compatibility?

Furthermore, the directional asymmetry observed in alignment quality hints at information-theoretic hierarchies between models. Some architectures may inherently discard information that others preserve, creating irreversible bottlenecks in cross-model translation. Understanding these information flows could guide the design of more robust and interpretable biometric systems.

Ultimately, this work reminds us that despite the apparent black-box nature of deep neural networks, the representations they learn often reflect fundamental structures of the physical world. The facial manifold, with its complex but constrained geometry, appears to guide independently trained models toward surprisingly compatible solutions, suggesting that in the high-dimensional spaces of deep learning, geometry may be more universal than we previously assumed.