MegaPortraits turns your photos into creepily accurate deepfake videos

Jul 25, 2022

Dunja Djudjic

Dunja Djudjic is a multi-talented artist based in Novi Sad, Serbia. With 15 years of experience as a photographer, she specializes in capturing the beauty of nature, travel, and fine art. In addition to her photography, Dunja also expresses her creativity through writing, embroidery, and jewelry making.

MegaPortraits turns your photos into creepily accurate deepfake videos

Jul 25, 2022

Dunja Djudjic

Dunja Djudjic is a multi-talented artist based in Novi Sad, Serbia. With 15 years of experience as a photographer, she specializes in capturing the beauty of nature, travel, and fine art. In addition to her photography, Dunja also expresses her creativity through writing, embroidery, and jewelry making.

Join the Discussion

Share on:

YouTube video

Deepfakes have been a thing for a while now, and they’ve been used for fun, useful, but also malicious purposes. Samsung has now introduced yet another deepfake tool that turns a single still image into a creepily accurate video.

Samsung’s Megapixel portraits, or MegaPortraits, is able to create high-resolution human avatars. All it needs is a single portrait, and it doesn’t even have to be a photo – you can also turn paintings into moving portraits, for some extra creepiness.

In their paper, researchers from Samsung Labs elaborate on the technology behind MegaPortraits, describing the training setup as “relatively standard.”

“We sample two random frames from our dataset at each step: the source frame and the driver frame. Our model imposes the motion of the driving frame (i.e., the head pose and the facial expression) onto the appearance of the source frame to produce an output image. The main learning signal is obtained from the training episodes where the source and the driver frames come from the same video, and hence our model’s prediction is trained to match the driver frame.”

Samsung AI Center in Mosci and the Skolkovo Institute of Science and Technology introduced similar technology back in 2019. In the latest work, they have made the technology more advanced. The team writes that they focused on a “challenging task of cross-driving synthesis.” This means that “the appearance of the driving image is substantially different from the animated
source image.”

While the facial features been improved, the researchers note that their method underperforms in terms of shoulders and clothing. This makes the resulting videos a bit jittery and less believable, but Samsung Labs says that improving this will be their focus in the future. I’d say that the hair also looks a bit weird when the photos and paintings begin to move, which is giving them away. Still, it’s hard not to think how much better deepfakes will become in the near future, and how easy they will be to believe even when you view them on a large screen where it’s easy to spot the imperfections with today’s deepfake videos.  So, don’t forget to keep your critical thinking on a high level as I believe we’ll see more and more believable deepfakes as time goes by.

[via PetaPixel]

Filed Under:

Tagged With:

Find this interesting? Share it with your friends!

Dunja Djudjic

Dunja Djudjic

Dunja Djudjic is a multi-talented artist based in Novi Sad, Serbia. With 15 years of experience as a photographer, she specializes in capturing the beauty of nature, travel, and fine art. In addition to her photography, Dunja also expresses her creativity through writing, embroidery, and jewelry making.

Join the Discussion

DIYP Comment Policy
Be nice, be on-topic, no personal information or flames.

Leave a Reply

Your email address will not be published. Required fields are marked *