- Tram Ho
Digital image rendering and editing can take hours and hours, but researchers from MIT and IBM want to change that.
Specifically, the researchers trained AI how to create photographs from zero and cleverly edit the objects present in the image. Although this AI will give advantages to artists and designers, it also shows us how the neural networks see and learn the context, and the team hopes to improve. This tool is a product that can detect fake or edited photos.
Named GANpaint Studio, this tool currently has a demo downloadable here. Instead of adding a tree yourself to the image, you can tell the tool to place the object in the position you want, and it will add the tree that best matches the context of the image. You can also delete objects, such as deleting a chair from the kitchen photo, for example. Although there are still many things to be done, the team hopes GANpaint Studio will one day be able to edit video clips, so that movie editors can use AI to add an important element. that was missed from the frame while shooting.
It is quite interesting that during the development of GANpain Studios, the researchers were surprised to discover that the system learned simple rules about the relationship between objects – like a wing. The door cannot be placed in the sky. Because GANpaint Studio uses GAN – a set of neural networks developed to compete with each other – it is forced to show its arguments to the decisions made, as to why it is impossible to place a bunch clouds on the grass … That information can help researchers better understand the neural networks about the context, and what they think in common sense.
Although GANpain Studios makes it easier to create fake photos, it also helps computer scientists discover fake photos. ” You need to know the opponent before you can fight it,” said study co-author Jun-Yan Zhu.
Source : Tri thức trẻ