multi-scene-video-assemblylisted

# Multi-Scene Assembly The instinct for a 20-second video is to generate a 20-second clip. It's the wrong move on every axis: it costs more, it drifts and morphs, it holds subject identity worse, and you can't fix one bad moment without re-rendering all of it. **Generate beats. Assemble in an editor.** That's how the output stops looking generated. ## Why short clips win | | One 20s generation | Four 5s clips, cut | |---|---|---| | Subject identity | Degrades over the clip | Holds within each | | Morphing and drift | Common | Rare | | Cost of one bad moment | Re-render everything | Re-render 5 seconds | | Iteration | Whole clip | Per beat | | Pacing control | Whatever the model does | Yours, in the edit | | Looks generated | Yes, usually | Much less | The last row is the one that matters commercially. Long single generations have a characteristic drifting quality that reads as AI immediately. Cuts read as editing. ## Storyboard first Decide the beats before generating anything. One motion per clip. ``` Beat 1 0-3s HOOK hands already mid-pour, close, fast Beat 2 3-7s PROBLEM the mess / the before state Beat 3 7-12s PRODUCT product in use, clear demonstration Beat 4 12-16s RESULT the after state, satisfying Beat 5 16-18s CTA product held to camera, room for text ``` Each beat is a separate generation with **one motion**. A beat that needs two motions is two beats. See `text-to-video-prompting`. ## Consistency across the beats