Artificially intelligent algorithms that create images are notoriously difficult to predict and control. But now researchers at Adobe have devised -controlled software that lets you transform the shape of objects in images, and adjust the lighting and perspective, with a few simple controls.
Given, say, a photograph of a car, you can use a combination of several control sliders in the program to adjust different components, including its shape and style, background, lighting conditions and colour. Dragging a slider can instantly change a convertible into a hatchback, for example.
The program also works for photographs and artwork of human faces, allowing you to convincingly alter features including age, wrinkles, hair colour, hairline, head angle and expression – ranging from angry to grinning.
Advertisement
Battling AIs
Erik Härkönen at Adobe Research in Finland and his colleagues built the program using existing algorithms called (GANs). GANs are comprised of two competing AIs. One, a generator, creates an image, while another, the discriminator, has to distinguish whether the image is real or fake. After several rounds, the generator becomes so skilled at producing realistic images that the discriminator can no longer tell the difference.
One of the AIs the researchers used was BigGAN, a DeepMind algorithm that has previously been trained on thousands of images associated with particular objects or animals, such as tools or birds.
The researchers analysed the existing GANs and broke down the way they each modify images into components, which they then labelled. Different component controls – such as orientation, time of day and degree of light reflection – could then be adjusted in combination with each other.
The researchers suggest that the program provides a simple yet powerful way to create images with existing AIs, without the computationally intensive need to train new algorithms for specific tasks.
Reference:
