- Tech

How to Make Realistic Photos With Artificial Intelligence

Yes, in a few (few) years, you find yourself in a magazine with the face of what appears to be a celebrity posing on the red carpet, which you do not finish placing (“Wait, is it Taylor Swift? Uhm, since here it looks more like Scarlett Johnson or whoever “), do not discard that in reality that person does not exist.

And, in the future, fashion publications can decorate their covers with photos of artificial models, created expressly by artificial intelligence from a database of faces of real celebrities.

Years back, a public project was developed in this regard, which used GANs (acronym in English ‘generative antagonistic network’) and the database to generate realistic photos of fake celebrities aggressively. This trend will soon, and if not yet be covered by the best photography books such as https://photolemur.com.

Here is the Tip

Machines that Learn to Generate Faces Playing Against Themselves

A GAN is a system based on the use of two neural networks: one that generates images and another which evaluates them until the first one is able to generate images that the second takes for good. It can be used to achieve more realistic photos or more attractive pictures.

In short, “the computer learns to generate these images by playing a game of cat and mouse against itself,” in the words of Jakko Lehtinen, one of the Nvidia researchers. This guarantees that humans will interpret most of them as real.

Specifically, the Nvidia researchers resorted to a progressive GAN, begins generating images of very low resolution and then gradually evolves towards higher resolutions, allowing their realism to improve more quickly, but also more stable. Result? Images of 1024 x 1024 pixels, crisp and convincing.

Researchers have recognized that “although the quality of our results is, in general, high if compared to other work done so far with GANs, there is a long way to real photorealism.”

Having said that, we feel that we can now achieve a convincing realism.”

Dare with the MIT Test

To prove to what extent researchers were right, the MIT website hosts an online test in which users are encouraged to try to recognize which faces are true or artificial (generated with Nvidia’s technology).

You can do it, but once you start, you will be shown 6 face images, which will flash for 5 seconds. After each image disappears, respond if you think the face was real or false. If you bet everything to your mere intuition and you have good luck, you can get to hit 50% but after several attempts, and knowing what to look for, you can achieve 100% success in two consecutive batches, as has happened to the editor of this article.

About Paul Petersen

Read All Posts By Paul Petersen