What sets Disney’s efforts apart from others’ attempts is its focus on megapixels. Traditionally, deepfake technology has focused on smooth facial transfers — that is, making a face look like a face on a face — rather than the nitty gritty details of an image. Models from DeepFakelab, for example, produced an image that was 256 x 256 pixels. Disney however, ramps that up to a 1024 x 1024 resolution. Not only does this give deepfakes a more realistic look, but it means the images will look better on bigger screens, which is evidentally what the House of Mouse has its targets on.
In a recently published paper, Disney Research Studios outlines how it’s utilized progressive algorithm training, stabilization technology and lighting effects to achieve thoroughly convincing face-swapping results. The results are a little unsettling, to say the least.
Disney's deepfake technology pumps out videos at high-resolution
The added megapixels help it create deepfakes at 1024 x 1024
It still has limitations before being used commercially
That resolution also means that the images will ultimately look better on bigger screens. As noted by The Verge, outside of the algorithm's high resolution, Disney's deepfake technology is fairly consistent with others of its kind. Though it can automatically adapt to different facial expressions, it has a difficult time with faces that aren't forward-looking and well lit.
0 Comments