Use this full prompt for the first controlled attempt
This prompt is designed for a fifteen-second vertical output using one consented portrait and the site’s supplied reference audio. It begins with format and camera because the exterior approach establishes the premise. It then protects identity before assigning the meal and reaction. The curtain is isolated as an ending with a start time, direction, deadline, and settled final state. Concrete negative constraints appear last so they do not interrupt the chronological sequence.
FORMAT AND CAMERA
Create one continuous 15-second, 9:16 vertical night video. The camera starts outside across a quiet street, facing a dark apartment facade with one warm illuminated window. The window occupies about one quarter of the opening frame. Make one slow, stable push toward it through second 9. Preserve the facade, glass, frame lines, room layout, and direction of travel. No cuts, orbit, pan, or jump zoom.
SUBJECT AND ACTION
Inside the same window, one consenting adult subject sits at a small table with one unbranded burger. Preserve the supplied face, hair, skin tone, clothing, body proportions, and identity. From 0–8 seconds, the subject looks down, calmly lifts the burger once, takes one small bite, and lowers it. At second 8, the action stops. The eyes turn to the camera, then the head turns slightly. Hold restrained surprised eye contact for about one second while the body remains seated.
CURTAIN ENDING
At second 11, the subject reaches to the existing curtain edge beside the chair and pulls it horizontally across the window with one hand. The fabric follows the hand naturally, fully covers the subject and glass by second 14, and remains closed through second 15.
VISUAL RULES
Photoreal low-light exposure, warm interior against deep navy exterior, believable reflections, subtle rain haze, realistic hands and fabric. One subject, one burger, one window, one camera move. No extra people, identity drift, face morph, duplicated fingers, floating food, new curtain, moving window frame, text, logo, watermark, or surveillance overlay.Run the baseline before replacing the burger, changing the room, or adding weather. The first result is a diagnostic. If the camera cuts inside, revise the camera block. If the subject looks up early, revise the starting gaze and reaction time. If the curtain appears from nowhere, revise the source composition and curtain block. Do not treat every failure as evidence that the entire prompt is wrong.
The production preset is 720p, 9:16, and fifteen seconds. Native audio is disabled because the provided fifteen-second audio file is sent as a reference. That keeps the sound input consistent while the visual prompt changes. The model may still interpret timing rather than obey it frame-perfectly, so judge whether the four story beats remain readable instead of requiring an exact frame number on the first attempt.
Check the clip against this second-by-second plan
Seconds 0–4: The facade and warm window establish location. The window remains small enough to feel observed from a distance, and the camera begins one stable forward move. The subject is visible but not yet performing for the lens. Reject a result that cuts indoors, changes buildings, or starts with direct eye contact because later prompt edits will not restore the intended reveal.
Seconds 4–8: The camera continues the same path while the subject completes one small food action. The head and eyes remain directed toward the meal. The burger should stay recognizable and unbranded, with no wrapper text or extra food appearing. If the hands are unstable, replace the bite with a lift and lower; the story needs an unaware diner more than it needs visible chewing.
Seconds 8–11: The food motion stops, eyes find the camera, and the head follows slightly. This is the reversal, so it must be visible without becoming theatrical. If the face morphs, reduce expression intensity and style language. If nothing changes, extend eye contact or add a small brow lift. Keep the person seated because standing introduces new body geometry just before the difficult curtain action.
Seconds 11–15: One hand reaches an existing fabric edge and draws it across the glass. Full coverage occurs by second fourteen, leaving a short closed-window hold. If the ending is incomplete, start at second ten or shorten the travel distance in the source image. If the glass itself slides, repeat that the frame remains fixed and the fabric—not the window—moves.
Change one visible variable while preserving the story
Swap the food
Replace “one unbranded burger” with one compact food that can be held predictably, such as a sandwich or slice of pizza. Keep the singular count, central hand position, and one lift. Soup, noodles, or a full table spread add utensils, flexible strands, and multiple objects. Those can become separate experiments after the camera and curtain already work.
Change the room
Replace the generic apartment with a sparse diner booth, studio kitchen, or late-night office, but preserve one warm window, a reachable curtain, and simple furniture. Do not change the room and facade in the same attempt. A different interior already asks the model to rebuild geometry behind the glass; changing the exterior simultaneously makes continuity errors harder to locate.
Adjust the reaction
For a calmer version, remove the brow lift and ask for a neutral pause before the curtain. For light comedy, add one brief double take without standing or dropping the food. Avoid strong fear, aggression, or a rush toward the camera. Besides increasing distortion, those choices shift the clip away from the small social reversal that makes the format readable.
Record each variation as a new version instead of silently overwriting the baseline. Compare only outputs made from the same source portrait and production settings. If a variation succeeds, you know which clause caused the visible change. If it fails, you can return to the known baseline. The broader prompt module library provides isolated camera, reaction, and curtain blocks for that process.
Match each common failure to a specific correction
Face changes during the approach: Use a sharper source with less occlusion, reduce extreme lighting adjectives, and repeat the exact identity attributes before the reaction. Keep the expression restrained. A rapidly changing camera angle creates harder identity work than a straight push, so remove orbiting or side movement.
Reflections hide the subject: Remove heavy rain and neon reflections, request subtle glass reflection only, and strengthen the interior-to-exterior exposure balance. The window must still read as glass, but clarity of the person is more important than a dramatic reflective surface.
Hands or food deform: Reduce the meal action to one lift and lower, keep the burger centered near the torso, and avoid switching hands. If necessary, leave the food on the table and begin the reaction from a resting pose. This sacrifices a bite but preserves the recognizable premise.
The curtain spawns or misses the window: Make the curtain visible in the input, place its edge within reach, name one hand and one direction, and start sooner. Do not ask for two panels. If geometry still fails, change the source composition before adding more text.
For the complete upload-to-review workflow, use the step-by-step tutorial. Return to the main guide for the narrative logic and disclosure reminder. KIE’s official documentation is the technical source for its multimodal reference fields; always recheck it before production deployment because provider contracts can change.
Technical reference: official KIE Seedance 2.0 Mini documentation.