Describe a single move the viewer can understand
The camera supplies the premise before the person does anything. It begins far enough away to show a dark facade and one warm window, then approaches without changing direction. If the prompt merely says “zoom into a window,” a generator may simulate a digital crop, jump into the room, or replace the facade with a close-up. Define where the camera is, how much of the frame the window occupies, when the move happens, and which geometry must remain unchanged.
CAMERA: One continuous exterior shot from across the street. Begin with the lit window occupying about one quarter of the vertical frame. Make one slow, mechanically stable push toward the window from 0–9 seconds. Preserve the building facade, window frame, reflections, perspective, and direction of travel. No cut, no orbit, no whip pan, no sudden digital zoom.“Mechanically stable” does not require a sterile result. Light rain, haze, distant traffic, or a tiny amount of natural camera drift can add atmosphere after the path works. Test those details separately. A handheld instruction often changes both the path and the lens behavior, which makes it hard to know why the window jumps. An orbit invites the model to invent the side of a building that never existed in the source. A drone description may send the camera upward instead of forward.
Frame size is another practical control. Starting with the window at about one quarter of the vertical frame gives enough exterior context while keeping the person visible. Starting with a tiny dot asks the model to invent facial detail during a huge scale change. Starting too close removes the voyeuristic setup. For a more intimate variation, adjust the window to one third of the frame but keep every other instruction. Compare the result before adding a different room, food, reaction, or ending.
Turn “looks surprised” into observable face and body cues
Emotion words are compact, but they leave the physical performance open. “Shocked” can become a scream, a jump, a dropped burger, or a distorted open mouth. The scene needs recognition rather than panic. Give the subject a stable starting behavior, delay awareness, then specify the smallest visible changes that communicate the shift. Eyes first and head second creates a readable moment without demanding a full-body turn inside a short interval.
REACTION: The subject does not acknowledge the camera during seconds 0–8. At second 8, the eating motion stops. The eyes move toward the lens first, followed by a small head turn. Hold restrained eye contact for about one second; brows lift slightly, mouth remains natural, body stays seated. No smile, scream, lunge, wave, or exaggerated horror expression.If eye contact happens from the opening frame, describe the starting gaze with an object and direction: eyes down at the burger, head angled slightly toward the plate, shoulders facing the table. Negative wording alone—“do not look at camera”—can be weaker than a positive pose the model can render. If the reaction is too subtle, lengthen the hold by half a second before making the expression larger. A longer readable pause usually preserves identity better than stronger facial deformation.
The food action should support the reaction rather than compete with it. One lift and lower is the safest. A small bite adds mouth and food contact, which may increase distortion. Chewing, wiping the mouth, switching hands, standing, and turning create too many simultaneous changes. Decide whether the essential cue is “caught while eating” or the exact bite. If the burger repeatedly fuses with the hand, keep it on the plate until the subject notices the camera. The narrative still works because the table and food establish the situation.
Give the curtain a source, direction, deadline, and final state
The curtain is the hardest prop because it combines a hand, flexible material, window geometry, and a full-frame transition. A vague closing instruction may make the glass turn opaque or cause fabric to appear in the center. Begin with a curtain that already exists beside the subject. Name the reachable edge, the hand that grips it, the horizontal direction of travel, and the second at which the action starts. Then state what the final frame contains.
CURTAIN: At second 11, the subject reaches to the existing curtain edge beside the chair, grips it with one visible hand, and pulls it horizontally across the glass in one smooth motion. The fabric follows the hand, covers the subject completely by second 14, and stays closed through the final frame. Do not create new fabric or move the window frame.Do not ask the model to draw two symmetrical panels unless the source image clearly contains both and the subject can reach them. One panel moving across the glass uses fewer interacting elements. If the hand duplicates, replace “reaches and grabs” with a starting pose in which the hand already rests near the curtain edge. If the fabric cuts through the body, move the subject farther from its path. These are composition repairs, not synonyms for the same prompt.
The deadline protects the ending. When the closing motion begins at second thirteen, the model may still be moving as the clip ends. Starting at eleven allows three seconds for travel and one second for a settled frame. If your audio has an obvious final beat, align the moment of full coverage with that beat rather than adding a new visual action. The site’s production preset uses the provided fifteen-second reference track and disables native generated audio, keeping the timing input consistent across attempts.
Combine modules carefully and keep a failure log
Start with the complete window burger prompt, then replace only the module related to your experiment. Save each prompt beside its output with a short note such as “face drift begins at 6.2 seconds” or “curtain appears from center.” A useful log records the source image, model, duration, aspect ratio, resolution, reference audio, prompt version, and first visible failure. Without that record, a lucky result cannot be reproduced and a bad result teaches very little.
Negative constraints belong at the end and should target observed errors. A generic wall of “no bad quality, no ugly, no artifacts” does not explain the scene. Prefer concrete exclusions: no second person, no extra burger, no logo, no hand duplication, no window-frame movement, no cut, no camera reversal. Add a new negative only after the corresponding problem appears. Too many prohibitions can dilute the positive sequence and make the output static.
Use the step-by-step creation guide for image preparation, production settings, review, consent, and disclosure. Return to the main Window Peeper guide for the scene’s four-beat structure. KIE documents reference images, videos, and audio as a multimodal input mode; it also says this mode should not be mixed with strict first-frame fields. That is why this library describes preservation goals without claiming pixel-identical opening frames.
Technical reference: KIE’s Seedance 2.0 Mini documentation. Prompt behavior remains probabilistic; the modules are production hypotheses to test, not guarantees.