Can we generate real faces from rPPG signals? Probably not

The potential of generating authentic human facial images from remote photo-plethysmography (rPPG) signals is a compelling idea, with significant implications for biometric authentication and human-computer interaction. This study explores it by using a large-scale dataset to train a diffusion-based generative model, leveraging rPPG signals extracted from facial videos. The initial training phase yields promising results, with the model demonstrating a capacity to synthesize facial likenesses that closely match the corresponding subjects in the training dataset. However, the performance notably falters during validation with an independent dataset, where a marked divergence between generated and actual faces becomes apparent. A subsequent human perception study corroborates this discrepancy. These observations suggest that rPPG signals alone may not be reliable for accurately generating realistic facial imagery.