We showed 100 people a real video and an AI one. More of them believed the AI.
Here is the test we ran, and we did not expect it to go this way.
Victor, our CEO, recorded himself talking to a camera. Then we took one frame out of that recording — a single still — plus the audio, and fed them into CraftStory. The model generated everything else from scratch: the gestures, the facial expressions, the lip sync.
That gave us two videos of the same person saying the same thing. One filmed. One generated from a photograph.
We showed each video to 100 people and asked one question: was this AI generated, or real?
The number that stuck with us was a different one: 14 people watched a real man filmed on a real camera and said it was definitely AI.
Two years, and most of it went into hands
Making one convincing second of a person is close to solved. Everybody can do that now. What took us two years was everything after those first few seconds, and the parts of a human body nobody thinks about until they look wrong.
Hands, mostly.
We shot our own training footage rather than scraping it, and we shot it at 120 frames per second. That sounds like overkill until you watch fingers. They move fast enough that ordinary 30 fps footage has motion blur baked into every take, and anything learned from blurred hands produces blurred hands forever. Hands are also the first thing a viewer notices when something is off — long before the mouth.
Nobody puts “we filmed at 120 fps” on a landing page. It is one of the reasons the videos hold up.
What still gives it away
When Victor showed the generated video on OpenCV Live, the host caught it. Not from the mouth, and not from the hands — from the eyes. The eye contact was too steady. The light on the face was too clean. Reflections behaved a little too well.
That is the honest state of it. The failure mode has moved from “the mouth is wrong” to “this person is lit better than anyone in real life.” We would rather say that out loud than pretend the problem is finished.
Watch the full talk
Victor covered all of this on OpenCV Live #218, with a live Q&A at the end.
CraftStory turns one photo into a video of that person talking, in 30+ languages. There is a free plan and it does not ask for a card. Last week we finished #12 of the day on Product Hunt.