A new technique that enables artificial intelligence (AI) to alter the age of subjects in photos without disturbing that subject’s identifying features has been developed by researchers from the New York University (NYU) Tandon School of Engineering.

Considered a major step forward from standard AI models that tend to make subjects look younger or older but without retaining their unique biometric identifiers, the researchers trained a latent diffusion model — a type of generative AI model — to understand how to conduct identity-retaining age transformation.

Source: NYU Tandon School of Engineering

To accomplish this, the research team reportedly overcame a common obstacle in the shape of having to assemble a large set of training data that consists of images that depict individuals over many years. Rather, the model was trained with a small set of images of a subject in addition to a separate set of images that featured captions to indicate the age category — child, teenager, young adult, middle-aged, elderly or old — of the subject depicted. The data set also featured images of celebrities captured throughout their lives.

The researchers explained that the model learned the subject’s unique identifiable biometric characteristics from the first set while the images featuring the age captions taught the AI model the relationship between images and age. As such, the model could eventually simulate aging or de-aging with researchers specifying a target age via a text prompt.

Further, a technique called "DreamBooth" was applied to edit human face images by gradually modifying them through a mixture of neural network components. This method involved the addition and removal of random variations or disturbances — otherwise known as “noise” — to images while taking into consideration the underlying data distribution.

Text prompts and class labels were used to direct the image generation process with the focus on retaining identity-specific details and overall image quality. Meanwhile, assorted loss functions were used to modify the neural network model, and the effectiveness of the approach was demonstrated via experiments on producing human face images featuring age-related alterations and contextual variations.

The technique was measured against other existing age-modification methods with researchers instructing 26 volunteers to match the generated image with an actual image of the subject, and with ArcFace, a facial recognition algorithm. The researchers suggested that their technique outperformed other methods.

An article detailing their method, Identity-Preserving Aging of Face Images via Latent Diffusion Models, appears in the journal arXiv.