One Prompt Rewrite Turned AI Video Characters From 'Moving' Into Actually 'Acting'
FinanceYF5 · x · 2026-10-02
The author shares a complete video-generation prompt showing that writing characters as personalities with an acting job—rather than just appearance plus motion—is what turns stiff output into genuine physical comedy.
Key techniques:
- Use the reference image only to lock identity and look (white hair, blue eyes, oversized sword, body-size ratio with a giant creature), explicitly stating it should not inherit the reference's fixed expression or poses;
- Give the character a personality (a cheerful, mischievous forest helper) and demand full-face-and-body acting, not just blinking and hair movement;
- Drive the interaction with specific lines, eyebrow raises, and emotional beats (Japanese dialogue included) instead of a bare action list;
- Add negative constraints: natural motion speed, camera following the action, no slow push-ins faking dynamism, no prolonged static straining shots;
- Start the action already evolving from the reference state (breath blowing her hair, bending flowers) so the scene doesn't sit frozen.
The full prompt is reusable in any image/video generation workflow.
Related event: New Prompting Techniques Turn AI Video Characters into Real Performers(2 posts)→
More from Multimodal
- Claude Opus 5.5 directed a 5.5-minute AI short film from one prompt in ComfyUI — Cheap_Credit_3957 · 2026-10-02
- Skipping perfect pixels: MiniMax H3 renders 15-second shots in 6 minutes on an RTX 3060 — Support_Marmoset · 2026-10-02
- MachGen pushes MiniMax H3 past its 15s cap with 30-second continuous video — MiniMax_AI · 2026-10-02
- NVIDIA and Waterloo release PixelUMM: an encoder-free multimodal model reading and writing raw pixels — CSProfKGD · 2026-10-02
- FLUX 3 Image launches with pixel-level control, multi-turn edits, 4K, and 10 reference images — Friendly-Fig-6015 · 2026-10-02
- Fable 5.5 outputs so good that even 'slop creators' are making insane videos — inductionheads · 2026-10-02