Animate from image-to-video (Seedream still → Seedance) instead of text-to-video so Seedance can match a reference picture and keep the character consistent across shots.
Only send “clean”/approved frames to Seedance for animation; filter out wrong-face/wrong-character stills first to avoid paying to animate bad frames.
Have your script model generate a per-shot prompt list (treat shot prompts as the deliverable), then feed those prompts into Seedream for still generation and Seedance for animation.
