New AI method generates realistic human faces
The researchers used a new approach to generative adversarial networks (GANs) — a class of AI algorithms used in unsupervised machine learning that pits two neural nets (a generator and a discriminator) against one another – to create the images. Traditional such networks however, say the researchers, offer very limited control over the generated images.
Instead, they propose an alternative “style-based” generator architecture that automatically learns to separate different aspects of the images. It starts from a learned constant input and adjusts the “style” – from “coarse” to “fine” – of the image at each convolution layer.
This architecture, say the researchers, leads to an automatically learned, unsupervised separation of high-level attributes – such as pose and identity when trained on human faces – and stochastic variation in the generated images (for example, freckles and hair), and enables intuitive, scale-specific control of the synthesis. After training, such attributes can be combined in any way desired.
“The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation,” say the researchers in a paper on their work.
In addition to being able to generate synthetic faces, the system can also modify features on real human images, such as age, hair, or skin color. It is also not limited to human faces – it can be trained to autonomously generate animals, cars, or even bedrooms.
“Based on both our results and parallel work by [other researchers],” say the researchers, “it is becoming clear that the traditional GAN generator architecture is in every way inferior to a style-based design.”
In addition to their new proposed generator architecture approach, the researchers introduced a new, highly varied and high-quality dataset of human faces. For more, see “A Style-Based Generator Architecture for Generative Adversarial Networks” (PDF).