Do you have a link to any of those AI images; I can’t even visualize how that’s done.
Just click on the word "Me^" at the top of the page.
“I can’t even visualize how that’s done”
The AI is trained to do this. Basically for a “deepfake” video they will start with two different videos (could be many more than two, but let’s keep it simple).
There is a base video that you want to alter. Say it’s a video of a woman eating an ice cream sunday, so we can keep it clean. Then you have a video of a celebrity doing something different, but with enough similar movements that the AI can collect images from the right angles and facial expressions. The AI will study that second video and then try to “paste” the face of the celebrity on to the face of the woman in the other video. The AI is doing this very roughly at first, frame by frame, trying to select the closest angle/facial expression from the second video to match the face of the woman in the first video. Keep in mind that the base AI has already been trained before this to recognize human faces and facial features like eyes, noses, mouths, etc.
After the AI makes that first pass, then a human being has to evaluate the results. Some frames will match very closely. Some not so much. Some will be horribly wrong. So the human basically rates those to give the AI feedback so it knows where it needs to go back and make changes. And that cycle continues until the human is happy with the results. And even after that, the human will probably go in frame by frame to do manual edits to clean up the final result.