Recognizing a specified subject in an image and automatically cropping the area can already be done with image editing software and a smartphone app. However, from here, AI that can be edited naturally, such as replacing objects, has emerged.
Researchers at MIT and IBM have used AI to create tools that can naturally insert or erase other objects in a photographic image. GANpaint Studio also has an online demo.
For AI learning, GAN, or hostile generation network, was used. For example, the generation network generates an image of a cat, and the identification network determines whether it is a cat or not. The generation network and the identification network learn to generate more precise images to be judged to be real, and the identification network learn to make a more accurate judgment.
The AI learned through GAN designates the type and size of objects such as trees, grass, and doors, and places them as if they would melt into the original image when you select the part you want to insert. However, it does not insert an image of an unnatural shape. For example, they don’t paint grass in the sky or clouds on the ground. Interestingly, this is not what the researchers expected, but the result of AI learning the rules for object placement by itself.
The research team revealed that using this technology there is also the possibility of identifying fake edited images. GAN Paint Studio demo can be found here , and related information can be found here .