What do you think to these portraits? Well lit, high quality, quite professional? What about if we told you they’re not photographs at all, that these people do not exist, and that the images are generated entirely by artificial intelligence?
NVIDIA, the company that released a surprisingly accurate Content-Aware Fill earlier this year, has now unleashed their next development. Their new paper details a generative adversarial network (GAN) which allows for easy customizing of the style of realistic faces.
As The Verge reports, GAN technology has existed for only four years. Although revolutionary at the time, below is what the results looked like. Needless to say, the attempts to form human faces looked more like surveillance video captures.
In the short time since then, the technology has improved so significantly that it can produce results that are indistinguishable from actual photographs. Researchers at NVIDIA are now able to sample source faces and replicate them onto destination faces, essentially taking samples of existing faces in order to blend and create images that have copied features but which look like new people altogether.
The researchers spent a week training the AI using GPUs. Here are some examples of the scarily realistic images the technology has created. It’s not just portraits; it’s been generating interior real estate photos, cars, and even cats.
What this means for the future of stock photography remains to be seen.
All images: Goodfellow et al; Karras, Laine, Aila / Nvidia.