A neural network to modify a photo like Photoshop

[ad_1]

Researchers affiliated with IBM and MIT, specializing in artificial intelligence, have developed a tool that transforms a photograph in the manner of Photoshop and its filling with content taken into account.

Researchers from IBM Research, the MIT CSAIL and MIT-IBM Watson AI Lab, teamed up to create GANPaint Studio. The tool allows adding or removing elements from a photograph while trying to respect the original image as well as possible. A function that is reminiscent of filling with content taken into account Adobe Photoshop editing software.

A small application usable online

Specializing in artificial intelligence, the small team used the neural network called GAN to create the software. Thanks to GANPaint Studio, it is possible to use a brush to select a part of the image to remove or to add an element. The small application is also available online with the researchers' images, but it is also possible to use your own photographs.

Thanks to the tool, it is possible to delete content, but also to add grass, bricks, trees or clouds. If GANPaint Studio is fast and sometimes proves surprisingly effective, it is above all for researchers to show a playful result and a practical potential on research that can be abstract.

Better understand the functioning of a neural network

While work had started well before 2014, the IBM and MIT teams wanted to better understand how the neural network works by answering a question: does the neural network work by learning composition or by memorizing patterns? pixels? From the first results obtained thanks to new research, it would seem that the tool is close to a composition method.

If this research cannot be used as it is outside of a small application that is fun for the general public, several perspectives are nevertheless possible. Of course, creating new, more efficient tools for image processing is one of them. However, research often aims to understand a particular mechanism before finding a practical application.

[ad_2]