Computers

Adobe advances AI-driven photo retouching

Adobe advances AI-driven photo retouching
Adobe has found an algorithm that adds the style of a reference photo to an original more convincingly than others. In this example, the original is at the top left, the reference photo is on the bottom left, and the result is on the right.
Adobe has found an algorithm that adds the style of a reference photo to an original more convincingly than others. In this example, the original is at the top left, the reference photo is on the bottom left, and the result is on the right. 
View 2 Images
Original and reference photo at top; result below
1/2
Original and reference photo at top; result below
Adobe has found an algorithm that adds the style of a reference photo to an original more convincingly than others. In this example, the original is at the top left, the reference photo is on the bottom left, and the result is on the right.
2/2
Adobe has found an algorithm that adds the style of a reference photo to an original more convincingly than others. In this example, the original is at the top left, the reference photo is on the bottom left, and the result is on the right. 

Automatic tools in Photoshop and Lightroom have come a long way, but so far, they've largely been useful for the nuts-and-bolts aspects of image manipulation. But according to a new research paper, Adobe is well on its way to automatizing the more artistic aspects of image retouching as well.

In the Deep Photo Style Transfer project, researchers from Adobe and Cornell University developed an algorithm that applies the style of a reference photo onto an original without many of the imperfections typically seen in AI-driven image manipulation.

For instance, an original photo of a daytime cityscape, paired with a night time reference photo of a different city setting, yields an end photo where night has fallen on the original city. The effect in the samples is photorealistic, without a false painterly effect, and seems relatively unafflicted with aberrations like cut-off buildings, inaccurate shadows, phony textures, distortions or artifacts.

Of course, we've only seen these results in a small set of low-resolution images, but they do seem promising.

Original and reference photo at top; result below
Original and reference photo at top; result below

Adobe has not confirmed if or when an algorithm like this could hit its software suite, but research like this seems to be a strong indication of the capabilities we could see in the future.

Source: Cornell University, via 9to5Mac

1 comment
1 comment
Daishi
I could see a ton of uses for this like re-theming houses and designs. Cool stuff.