New AI Can Re-Create Aesthetic in Order to Merge Two Different Photos Together

New AI Can Re-Create Aesthetic in Order to Merge Two Different Photos Together

A new AI tool can take the “visual style” or aesthetic of one photo, and astonishingly overlay it onto a second photo to look like a real picture.

Photo technology is advancing at an alarming rate. Who can forget the AI Google released at the end of last year, which they claim can “score” photos based on their technical quality and aesthetic? NVIDIA have never shied away from AI technology (they’re the company behind the software that creates new faces from celebrity portraits), and now have announced their latest project, FastPhotoStyle.

The AI software takes a “style photo” (let’s say a snowy scene) and merges it together with a “content photo” (perhaps, a photo of a house) to create a new image. In this example, it would become a photo of a snowy house. Its creators described the process as Photorealistic image style transfer[ing] algorithms [aimed] at stylizing a content photo using the style of a reference photo.”

With significant competition in the market already, the developers insist the most important aspect of their new product was that the final results must remain “photorealistic.”

One of its inventors, Ming-Yu Liu, insisted people were twice as likely to select the AI’s renderings as their favourite, when tested against manual photos taken by humans. The tool is also said to work “60 times faster” than some of the industry’s current systems.

Log in or register to post comments

3 Comments

Matthew Saville's picture

Incredibly impressive, on the one hand, but also this is yet another reason for artists to take more pride in the disclosure of their artistic mediums, instead of clinging desperately to the title "photographer" as if it's the only title that offers a path to prestige. Hopefully, we'll see many folks proud of being a "digital artist" (or whatever term catches on)

I've always been impressed by this kind of work, and I have tons of respect for the artists who master these techniques, whether they're making wall art or SFX in movies. I'm very glad that the era of bad CGI is behind us! (And Lucas / Spielberg / Cameron etc. would have killed to have these AI tools at their fingertips 30 years ago...)

However, as an outdoor photographer, I've always placed extra value on the "realness" that a photograph CAN hold. Because, like it or not, there IS an extra "wow factor" inherent in knowing that the light, the subjects, their scale and juxtaposition and timing, everything, ...all came together "for real".

We can argue "all photos are edited, so who cares?" all day long, but here's the bottom line- all photos are either more realistic, or less realistic, WRT laws of physics, optics, weather. etc. ...and again, like it or not, viewers still place at least some value on "real".

All this to say, if your digital art is entirely fabricated, by your careful brush strokes or your AI software, then the image should stand on its own two feet as a product of your imagination (and technical prowess) as an artist, ...NOT as a "real" moment. Often times, "bad composite" artwork just doesn't stand on its own two feet, and everybody flies into a defensive frenzy/rant when someone says "ehh, this looks dumb". The truth is, it's not a dismissal of the art form itself, it is simply a matter of the art form being in its infancy, as photography itself was nearly 200 years ago. (And it took a while for photography itself to be considered art, as well!)

Spy Black's picture

Companies like Getty are going to have a field day with this, ripping people's work off, mutating it, and selling it as their own.

I see a huge advantage for real estate photographers who want to apply a sunny scene to a very overcast day. We can all add a blue sky with cheerful puffy clouds, but then we have to work on the rest of the photo to try to approximate a sunny overall color. I think it would be great to just put it through some automated program, even if we had to do some final tweaks to taste after the result was generated.