Researchers develop model to maintain identity in generated faces

Imperial College London researchers have developed Arc2Face, an advanced computer model that generates realistic images of human faces. What sets it apart is its reliance on identity embeddings, which enables the model to generate varied facial images while maintaining identity consistency. However, the current version of the model can only create images of one person at a time based on the provided ID embedding.

The proposed Arc2Face model builds upon the Stable Diffusion text-guided diffusion model. It is conditioned on ID embeddings extracted from a pre-trained ArcFace network, without relying on text embeddings.

The efficacy of this computer model is demonstrated by the authors, who trained a facial recognition model using synthetic images generated by their model. The resulting performance surpasses that of present synthetic datasets.

While generative models such as StyleGAN are capable of producing facial images, they often are impacted when attempting to maintain consistent identity during manipulation. Similarly, text-guided diffusion models like DALL-E 2 encounter issues when trying to separate identity from textual descriptions.

“We meticulously study the problem of high-resolution facial image synthesis conditioned on ID-embeddings and propose a large-scale foundation model. Developing such a model poses a significant challenge due to the limited availability of high-quality facial image databases,” researchers explain.

Research has shown that relying solely on single-image per-person datasets such as FFHQ is not enough to create a reliable computer model. To address this, the team behind the Arc2Face model utilized the comprehensive WebFace42M database, which features substantial intra-class variability, as the foundation for their work.

However, despite the benefits of large-scale datasets, the scientists note that current limitations still restrict them to tightly cropped facial regions for face recognition training. As such, there is a need to incorporate a wider range of diverse and unconstrained facial images to enhance the model’s capacity to generate complete face images.

“We perform extensive quantitative comparisons to evaluate the performance of recent ID-conditioned models in generating both diverse and faithful images of a subject,” the researchers say.

In the paper, the authors explore the use of the model for producing synthetic images that can help in training face recognition models. By sampling identity vectors from the distribution of ArcFace embeddings and ensuring sufficient diversity among the synthetic identities, they establish a dataset that serves as a resource for training face recognition models.

To ensure a diverse range of subjects with unique appearances, synthetic identities are selected based on a similarity threshold. The effectiveness of facial recognition models trained on these synthetic images is then compared to models trained on real-world datasets and other synthetic datasets generated through various methods.

According to the analysis, the biometric models that were trained using the Arc2Face-generated synthetic data show high performance across several evaluation benchmarks. In comparison to other synthetic datasets such as SynFace, DigiFace, and DC-Face, the FR models trained with Arc2Face-generated synthetic data showed superior results.

“Our experiments demonstrate its ability to faithfully reproduce the facial ID of any individual, generating highly realistic images with a greater degree of similarity compared to any existing method, all while preserving diversity in the output,” the researchers conclude.

Article Topics

Arc2Face  |  biometrics  |  biometrics research  |  facial recognition  |  synthetic faces  |  training