Production Pipeline

Ten Ideas to a 1:24 Master

How a character and a world get built once, then held across eight shots and eighty-four seconds without drifting - from a wide net of story ideas down to a graded master.

IDEATION · CHATGPT ASTRA SCRIPT + CHARACTERS · CLAUDE OPUS CHARACTER DESIGN · GROK IMAGINE REFERENCE SHEETS · NANO BANANA VOICE · VIBEVOICE, LOCAL COMFYUI GENERATION · MINIMAX H3 FLA2VIDEO, COMFY CLOUD ASSEMBLY · DAVINCI RESOLVE 10 story ideas free brainstorm pass, no format constraints yet Script + character descriptions Claude Opus - picks one idea, writes the 1:24 script and cast Grok Imagine - iterate the look ~10–15 generations, chasing one face CHARACTER DESIGN LOCKED Nano Banana - reference sheet front / three-quarter / profile, one character sheet from the locked design Nano Banana - first & last frame one pair per shot, ×8 shots - the bookends H3 will interpolate between VibeVoice - voice sample(s) generated locally, per character MiniMax H3 - FLA2Video first frame + last frame + voice → each shot 8 SHOTS × 3 VERSIONS = 24 RENDERS 16:9 · 1280×720 · under an hour, total render time Best take picked, per shot Import 8 selected clips + voice Edit and cut to final runtime - 1:24 MASTER EXPORT reference sheet first / last frame ×8 voice
The route we ran, start to finish. Grok Imagine's ~10–15 passes bought the locked character design that every downstream step depends on - reference sheet, frame pairs, and all eight generations trace back to it. Gold marks the batch and export milestones. Each of the 8 shots got 3 H3 renders (24 total, under an hour of compute); the best take per shot went into the Resolve cut.
asset / step hand-off batch or export milestone
ChatGPT Astra
Open-ended brainstorm - 10 story ideas before one gets picked.
Claude Opus
Writes the 1:24 script and the character descriptions the design work builds from.
Grok Imagine
Iterates the character's look - ~10–15 generations to lock the design.
Nano Banana
Character reference sheet, plus one first/last frame pair per shot (×8) as H3's I2V bookends.
VibeVoice
Local - zero-shot voice sample(s) for the character.
MiniMax H3 · FLA2Video
First + last frame + voice → clip. 8 shots × 3 versions, 16:9 1280×720, < 1 hr total render.
DaVinci Resolve
Import the best take per shot, cut to 1:24, master export.