The “Wish You Were Here” bots tricked humans trying to spot the fake up to 47 per cent of the time – could this tech lead to more claims of doctored ‘deepfake’ images?
Facebook is developing ways to insert computer-generated likenesses of people into photographs in an artificial intelligence project called “Wish You Were Here”.
The social network wants to make it easier for users to “blend” people into their images even if they were nowhere near the original subjects at the time. In a study, AI software created a map around the person due for insertion into a scene, then approximated the pose of the people in the photo before adding the third party’s approximated likeness into the shot. The research, using thousands of images, was aimed at “finding a way to add a person, without disturbing the persons already there”. It tricked volunteers trying to spot the fake human nearly 43 per cent of the time on average, depending on group size, however, the image quality of people created by the AI was decidedly mixed. What the AI sees when scanning the target person, right, and finding the right spot to spawn them into the other scene, shown by the blue box (Oran Gafni / Lior Wolf / Facebook). Fake images were created using three AI processes called Essence Generation, Multi-Conditioning Rendering and Face Refinement for “context-aware human generation”. These generated “semantic maps” to try and maintain the existing scene’s context before inserting scans of the new person’s hair face, torso, clothes and shoes as a “photorealistic” cut-out. Researchers said they found a “convincing ability” of the AI to add people into existing photos “while preserving the overall image quality”, but the software struggled recreating some human features, such as hairstyles.
Also, people added into pictures sometimes suffered “missing parts” due to the scene’s composition. Jagged edges remained in some computer generations – but it fooled many human checkers (Oran Gafni / Lior Wolf / Facebook). The AI was trained on more than 20,000 sample photographs from an open-source gallery, resulting in nearly 53,600 example images for analysis. Human volunteers were then asked to see if they could find the artificially-added people in group shots. They managed to spot the computer-generated people between 28 per cent and 47 per cent of the time, depending on the group’s size. The AI struggled with hair and matching up some body parts to scenes (Oran Gafni / Lior Wolf / Facebook). The research paper, due to be presented at the US Conference on Computer Vision and Pattern Recognition conference, is a collaboration between Facebook AI Research and Tel-Aviv University. Co-authors Oran Gafni, an AI research engineer for Facebook, and Professor Lior Wolf, of a faculty member of the university’s School of Computer Science, described the technique as a “novel method for inserting objects, specifically humans, into existing images” while “respecting the semantic context of the scene”. The study says: “Unlike other applications, such as face swapping, our work is far less limited in the class of objects. “In an extensive set of experiments, we demonstrate that the first of our networks can create poses that are indistinguishable from real poses, despite the need to take into account the social interactions in the scene.” Facebook’s latest project follows Google’s work using AI methods to insert objects such as vehicles into images by predicting scale, shape and location.