MIT creates AI soon to be able to create (and detect) retouched images

GANpaint Studio

Researchers at MIT and IBM have developed an AI that can generate, delete, or modify objects on an existing image. Ultimately, the team behind this tool wants to improve it so that it can detect retouched or created images using AI. 

This initiative, called GANpaint Studio, could eventually be very useful for artists and designers. With a few clicks, the tool allows us to add elements (vegetation, clouds, door, bricks ...) to an existing image and to delete others in a dynamic way. Available as a demo at this address, GANpaint Studio is still far from finalized (currently rendering seems very random from one photo to another),

GANpaint Studio, a real test tube for neural networks

Powered by GANs ( Generative Adversarial Networks ), algorithms developed to compete with each other to acquire skills autonomously, the tool already manages to impress its creators. Spontaneously, the system has indeed managed to learn simple rules governing the relationships maintained between the different elements of a given scene. For example, a door has nothing to do in the sky. 

GANs oblige, the algorithms that propel GANpaint Studio must provide researchers with the reasoning that justifies their decision-making. These observations could, among other things, enable MIT and IBM teams to better understand how neural networks manage to learn.what we consider as humans as common sense, explains Engadget.

Possible applications in the field of cinema (but not only)

In practice, this experiment could lead to more advanced software capable, for example, of modifying images within videos or even films. If a crucial accessory to the plot has been forgotten from a scene, it could be added later via an AI. 

Conversely, objects that have nothing to do in a scene and that can not be cut off during editing could be deleted.