For the past two weeks my friend Gry and I have been building a pop singer who does not exist. Her name is Mila Zima. One song is out, and on the same page as the video we listed the eleven things still wrong with it.
The song is Porcelain. Three minutes 58, soul-pop with a rock edge. We drew her face once in November and locked it, and every frame gets measured against that face before anyone sees it. Her voice was generated once and kept, so she sings everything with it. A person still picks the take by ear.
The first video took 13 hours and 45 minutes. We generated 175 takes and kept 56 shots. Those takes cost $177.25, and only $55.12 of that is on screen. Two days later we released an acoustic version of the same song. That one has twelve known flaws.
Most generated acts get found out after the audience works it out. Deezer and Ipsos asked 9,000 people this month, and 80% said AI music should be clearly labelled. We started from the label. Her site says, before anything else, that her face, her voice and the rooms are generated, and that the decisions are ours.
The first video has about 4,600 views in four days. That is not a chart story. The second film is being shot now. Every song gets the same page: the song, the film, what it cost, and what is still broken, so the next one starts from that list.
The full record is at https://milazima.com and the stage by stage account is at https://milazima.com/stages/