As you'll get to see in the video below, the software can work on the important bits, as well as on the fine details and there are three layers of refinement when combining the source image with the others.
The tech company released a paper on the matter, which you can find in the attached PDF, with this describing the endeavor. The company's researchers trained AI for one week, relying on eight Tesla V100 GPUs.
NVIDIA introduced the idea of GAN in a paper published back in 2014, but the human faces rendered back then (more on this below) were nothing like the realistic images we have now.
If we zoom in on these contraptions, some of the cars seem to have "please kill me" written all over them, while others appear to be legit. It looks like Asian designs are the best, while we also found a few American cars that were properly represented. German toys? Not so much, since the Porsche and the Beetle we have here look like they sat under the sun for way too long.
Oh, and if you think the cars look odd, you should see the people. Unsurprisingly, the software's most accomplished creation seems to be a set of human portraits, the kind you and I could have serious trouble marking as fake.
We've added an image of the portraits to the gallery above and if you've made it this far without getting dizzy, you should probably be OK with the cats and the home interiors the AI has come up with.
The bottom line is that we have to prepare ourselves for the risks and challenges associated with this kind of technology. New areas of interest related to this will undoubtedly develop in the future, as we'll all have to learn to tell the real images and videos apart from the faux ones.