How We Made the World’s First Transparent Deepfake
One frigid morning in Amsterdam, I had the privilege of witnessing the creation of the first transparent deepfake video. The creative process began at Amsterdam Studios in the Netherlands, where our creative partner, Revel.ai, is headquartered.
Revel.ai is a digital artist agency and industry leader in hyperrealistic, synthetic content creation using artificial intelligence or AI. They work with some of the world’s biggest brands in the entertainment industry to create award-winning visuals for films, series, and commercials. For instance, they were behind the viral Morgan Freeman deepfake. Upon entering the studio, it was clear that the setup process was extensive, as there was a considerable amount of equipment arranged for the shoot in the studio. An array of massive lights towered over the Revel.ai crew. The beaming lights encircled a 19-camera rig that required several hours to meticulously set up. The cameras were equidistant from one another and carefully mounted with lenses focused toward the seat in which the subject would be seated. Their distance was measured using bright tape that started at a single point, the subject’s seat, and spread out to form a half circle. It was set up in this way to ensure that all angles of the face would be recorded.
Nina Schick was to be the subject of the trailblazing deepfake. Nina Schick is an author, entrepreneur, advisor, and public speaker specializing in Generative AI. She is dedicated to making AI accessible and is one of the first experts in the field. She is the Founder of Tamang Ventures, an advisory and creative firm focused on Gen AI, and the author of the seminal best-seller, Deepfakes: The Coming Infocalypse, making her the ideal fit for this revolutionary deepfake video. I was struck by Nina’s passion and excitement for AI. In our discussion, while filming, Nina emphasized that AI is one of the most powerful technologies of our time. She described it as a tipping point that will impact humanity beyond comprehension.
Bob de Jong, Co-Founder and Creative Director of Revel.ai recorded Nina’s facial movements to capture various emotions. She expressed a wide range of emotions, from happiness to sadness to anger. She was prompted to exaggerate her face by laughing loudly and dramatically furrowing her brow to capture her feelings authentically. Nina’s expressive face allowed her to make facial micro-movements. She conveyed the full range of facial mannerisms with relative ease. It was also necessary to capture her face from all angles. She was instructed to shift her head from side to side and up and down. To capture a full range of facial expressions, Nina was also asked to speak, making sounds that allowed her mouth to contort. This enabled Revel.ai to accurately record various movements, such as smiles, frowns, and other subtle gestures in order to create an accurate representation of her facial poses. This footage would give a complete picture of Nina’s emotions allowing the AI models to recognize each emotion with accuracy and ease.
The initial shoot was complete but only the beginning of the process for Revel.ai — in the coming weeks, they would capture footage of the model in the video, record Nina’s voiceover, and train AI algorithms. After Revel.ai trained an incredibly life-like AI model of Nina until the output was indistinguishable from the real Nina, Truepic cryptographically signed and timestamped the video. The output was signed as an AI generation with Truepic’s transparency tools, including the named creators. The video was then published with transparent details. The breakthrough video was marked with a tamper-evident seal using Truepic’s transparency tools, which are compliant with the open content provenance standard developed by the C2PA. Truepic’s signing solution allows digital content to transparently display where it came from, eliminating any confusion resulting from AI-generated content, showcasing a future where all content on the internet can be transparent.
Watch the Video: Mirror of Reflection
