First, the software used a "conditioning" network to compare the low-res images to existing high-res photos in its database. Did you then think to yourself "that's so dumb; imaging doesn't work like that"?
The result was a series of 64×64 pixel images that add realistic detail to the original 8×8 images and are surprisingly similar to the 64×64 originals used to assess the tool's output.
The first part, called the conditioning network, attempts to map the base 8 x 8 image against other high-resolution images.
The software starts with pixelated, nearly unrecognizable source images, and adds more detail to it. This revolutionary invention from Google Brain can actually provide improvements on forensics.
The trope, most especially seen on criminal-themed shows, sees law enforcement or detectives unabashedly bark "Can you enhance that?" when referring to an obviously grainy security footage that simply can't be enhanced in reality. This network gathers a lot of high-resolution images to match the main source.
Now the "Pixel Recursive Super Resolution" model is trained up for working on cropped celeb faces and hotel bedrooms.
Gabriel Jesus inspires Man City - but what now for Aguero?
I think that they have a great squad, a lot of fantastic players, and they played a really good game against West Ham . But City boss Guardiola is adamant the 10-point gap between his fifth-placed team and Chelsea is too much to overcome.
Then using a second network, which the researchers call prior network, it added the high-res details to sharpen the blurred photos.
Ars Technica offered a simplified explanation of the process: "For example, if there's a brown pixel towards the top of the image, the prior network might identify that as an eyebrow: so, when the image is scaled up, it might fill in the gaps with an eyebrow-shaped collection of brown pixels". The end result usually contains the plausible addition of new details. When human observers were shown a real high-resolution celebrity face vs. the upscaled computed image, they were fooled 10 percent of the time (50 percent would be a flawless score).
The results are even better without the complexities of the human face to contend with - it fooled people with a success rate of about 28 percent when the photos were of bedroom interiors instead. A ideal score is 50 percent, so Google Brain fared okay, especially when considering that normal bicubic scaling didn't fool any observers.
Currently, the AI creations are the machines best guesses rather than accurate portrayals.
Anyone who has followed crime procedural TV shows, particularly the likes of CSI, will probably be familiar with the "zoom in, enhance" method of pulling evidence out of thin air, so to speak. The results might not be flawless as it is in Hollywood, but it could help to narrow down the list of suspects.