A virtual-singer IP we run ourselves. Her look, her sound and the visuals for each song are written down as rules first, then handed to AI: the songs, stills and video clips are generated, while people choose, edit and sign off. The first EP has six songs, and the first full music video is public on YouTube.
The challenge挑戰
The hardest part of AI imagery is keeping a character looking like herself.
Extra fingers, stray people and garbled text are common in AI stills; checking each one by hand is too slow.
A three-minute video needs 30-plus shots of two to five seconds, each landing on the right lyric.
What we built做法
A creative-director AI drafts the shot list against the lyric timeline; a reviewer AI checks it against six rules and sends it back, up to three rounds, before a person steps in. Gemini then checks every still for fingers, head count, text, proportions, light and more, discards anything under 5 out of 10, and a person makes the final pick.
Seven fixed features of her look and a wardrobe per song, plus rules for the sound: a late-night empty room, piano, synths and electronic drums only.
Every still carries two face references, a scene reference and a fixed description; each shot gets four candidates and one is chosen.
Speech recognition and beat analysis turn each song into a lyric timeline; chosen stills become eight-second clips cut to the lyrics, while the vertical shorts are animated lyric videos rendered in code.
One tool posts to Instagram, Facebook and YouTube together, after a person confirms.
Watch影片
Results數字
songs on the first EP
full music video
vertical shorts
rounds of AI storyboard review, at most
Built withStack
AI Film & ContentAI 影音製作
Start with a 30-minute call: we'll understand your situation, then propose the smallest useful version.