An exciting short demonstration video of a “neural photo editor” is doing the rounds on photography blogs and it might herald a future of Photoshop in which massive changes can be made with relative ease.
The Neural Photo Editor was developed by Edinburgh Centre for Robotics researcher Andrew Brock and uses machine learning to predict the way in which a user is planning to retouch a photo based on a “contextual paintbrush”.
It’s a relatively simple process in which users select a color and the system analyses the color in the context of the image in order to produce an intelligent output. Using a dark colour on a subject’s forehead can, for example, produce bangs or if you use the brush on the subject’s hair, it can change the colour of the hair.
Brock says in the video demonstration that the system works reasonably well for system-generated images but not for other photos. This can, to some extent, be circumvented by using an “adversarial network” – an algorithm which can compare changes between original and modified photos in order to apply changes naturally.
The system can be downloaded from GitHub but is still very much in the early stages and can sometimes create, in Brock’s own words, “bizarro” results.