It’s not yet over: AI tools continue to improve their ability to manipulate images. For now, the latest example is a research document, but it’s a very impressive tool that lets users drag elements to change their appearance.
Take a look below at some examples to see what this system is capable of.
You can change the size of a car, or turn a smile into frown, with just a click and drag. You can also rotate the subject in a photo as if it was a 3D image, changing, for instance, the direction that someone is facing. In one demo, the user can adjust the reflections of a lake or the height of a mountain range by clicking a few times,
The videos are from the team’s homepage. However, this site has crashed due to the traffic that Twitter sends to it ( mostly by user @_akhaliq who highlights interesting AI papers. It is worth following if you have an interest in AI).
The team behind this project notes that the real interest in the work isn’t the image manipulation itself but rather the user interface. Since a long time, we’ve used AI tools such as GANs to create realistic images. However, most methods are not flexible and precise. If you tell an AI image creator to “make a photo of a lion sauntering through the Savannah”, you will get a picture, but not the exact pose you need.
DragGAN is a model that offers a solution. The interface is the same as the traditional image warping but instead of smudging or mucking up existing pixels, this model creates the subject from scratch. The researchers state: “[O]ur method can create a sense of occlusion, such as the teeth in a lion’s mouth. It can also deform the object according to its rigidity, such as the bending leg of a horse.”
This is a demo, so it’s not possible to fully evaluate the technology. How realistic are the final images? The low-resolution videos are difficult to judge. It’s a good example of how to make image manipulation easier.