How many total generations did the video take?

38 shots across five worlds, averaging roughly 6-7 regenerations per shot before one cleared review, for just over 260 total generations. The jungle world alone - where the dress had to visibly transform into greenery without breaking the actor's face or silhouette - accounted for nearly a quarter of that count.

How long did the production take from brief to delivered video?

Twelve business days on the Signature package: two days to lock the storyboard and character bible before any narrative shot was generated, roughly seven days of shot generation and review across the five worlds, and the remainder for color grade, sound design, and two revision rounds.

AI69 / Field Notes 2026-09-04 8 min read

Case Study: Building a Five-World AI Music Video Without Losing the Character (Alive Tonight)

Q: How many worlds does the Alive Tonight video move through, and why does that make consistency harder?

Five: an airport terminal, a desert at sunset, freefall through clouds, a jungle where the dress grows into living greenery, and a city of spiral towers. Each world change resets lighting, environment, and often costume at the same time, which is the exact moment a locked reference frame is most likely to drift - so every transition got its own regeneration budget instead of sharing one across the video.

AI69 built AGA BORYN's Alive Tonight across five worlds and 38 shots. Here is the character bible, the regeneration count, and the exact transition that ate a quarter of the budget.

"Alive Tonight," the AI69 video for AGA BORYN, moves through five worlds and 38 shots and took just over 260 regenerations and 12 production days to hold one consistent character across all of it. That is the case study: not the finished 3-minute-40-second clip, but the character bible, the shot-by-shot review, and the one world transition that alone accounted for nearly a quarter of the total regeneration count.

What was the creative brief?

The brief called for a single narrative arc, not five disconnected visuals: an artist moving through an airport terminal, a desert at sunset, freefall through open clouds, a jungle where her dress grows into living greenery, and a city of spiral towers - one continuous story, one recognizable face throughout. The five-world structure was the point of the video creatively. It was also the single biggest technical risk, because every world change resets lighting, environment, and often costume at the same moment a locked reference is most likely to drift.

What got built before a single narrative shot was generated?

Two days went into a full shot-by-shot storyboard mapping all 38 shots to the track's structure, plus a character bible: one locked reference image for the artist's face and body, and five additional reference images - one per world - covering each costume variation, including the jungle costume in both its "before" and "mid-transformation" states. Nothing narrative got generated against a blank prompt; every shot referenced back to whichever bible image matched its world and costume state.

Which world took the most regenerations, and why?

The jungle sequence, where the dress visibly grows into living greenery mid-shot, accounted for close to a quarter of the video's total regeneration count on its own. That world asked the model to hold three things steady at once through a visual transformation - the actor's face, the silhouette of the dress, and a physically plausible growth of foliage across the fabric - and any take where the transformation read as a costume swap instead of a continuous change got rejected. The freefall-through-clouds sequence was the second-hardest, purely on motion: a body in freefall reads as physically wrong the moment limb position or camera distance breaks continuity between shots.

What did quality control actually check before a shot shipped?

Every candidate got checked against the matching bible reference on three points: does the face still match, does the costume state match what the story needs at that moment (base outfit, transitional, or fully transformed), and does the motion between this shot and the one before it read as continuous rather than a jump cut in body position. A shot failing any one of the three went back into the regeneration queue - across 38 shots, that review process is where most of the 12-day timeline actually went, not the initial generation of any single take.

What did the final numbers look like?

38 shots, five worlds, five locked costume references plus one base identity reference, just over 260 total generations averaging 6-7 per shot, two revision rounds, and 12 business days from brief to delivered video on the Signature package. None of those numbers show up in the finished 3:40 clip - what shows up is an artist the audience recognizes in a desert, in freefall, and in a jungle where her dress is turning into leaves, without ever looking like a different person.