The shot in your head is 16:9. The render is not.
You picture the scene perfectly: two leads in a glass office lobby on a rainy night, city lights behind them, tension hanging in the air between one line of dialogue and the next. Then the render comes back in 9:16 and the picture falls apart. The male lead's forehead is cropped out of the frame. The female lead is a small figure stranded at the bottom. The lobby that was supposed to sell "rainy-night office romance" is a grey smear behind them.
This is the most common framing failure in vertical microdrama, and it is not a model problem. It is a prompt problem. A 9:16 frame is not a cropped 16:9 frame. It is a tall, narrow frame with its own rules — and if your prompt still describes a wide cinematic shot, the model will do exactly what you asked: cram a wide shot into a tall frame, and cut whatever does not fit.
Once you write framing into the prompt the way the model actually reads it — camera height, distance, headroom, and where each character stands in the vertical space — the same scene renders cleanly the first time. Here is how.
1. The tall frame changes what matters
In 16:9, the eye travels left to right. In 9:16, it travels top to bottom, and the frame punishes anything that spreads sideways. Three rules follow from that:
- Headroom is a number, not a feeling. "Medium shot" tells the model almost nothing about where the top of the head should sit. "Headroom of one hand's width above the hair" does. Name the gap and the model holds it.
- Stack characters vertically, not side by side. Two leads standing shoulder to shoulder need a wide frame. In 9:16, place one slightly closer to camera and one half a step behind, or seat one and stand the other. Depth replaces width.
- One background element per shot. A lobby, a skyline, a rain-streaked window and a reception desk cannot all share a 9:16 frame. Pick the one that tells the story — usually the window or the lights — and let the rest fall out of frame.
2. Five framing prompts that work in office romance
These are patterns, not magic words. Swap in your own characters and lock cards, keep the camera language identical, and the frame stays stable across the scene.
The lobby two-shot. "Vertical 9:16, medium two-shot in a glass office lobby at night, woman standing left foreground, man half a step behind right, both fully in frame from head to waist, rain-streaked glass and city lights behind them, camera at chest height, slight headroom above both." The "head to waist" and "slight headroom" clauses are what keep both faces in the frame.
Over-the-shoulder at the desk. "Vertical 9:16, over-the-shoulder shot, man's shoulder and back of head soft in the lower foreground, woman in sharp focus across the desk, camera at seated eye level, her face in the upper third of the frame." Putting the face in the upper third uses the tall frame instead of fighting it.
The elevator close-up. "Vertical 9:16, tight close-up, two faces in one frame, foreheads near the top edge with a finger's width of headroom, elevator mirror and warm ceiling light behind, camera at eye level, no tilt." Elevators are the easiest office-romance location in vertical — the space is already tall and narrow.
The rainy window medium shot. "Vertical 9:16, medium shot, woman standing at a floor-to-ceiling window, back three-quarters to camera, city lights and rain on the glass filling the upper half, camera at shoulder height, her silhouette fully in frame." The window fills the top half; the character anchors the bottom half. Nothing is cropped because nothing competes.
The walking-and-talking corridor shot. "Vertical 9:16, medium tracking shot, two people walking side by side down an office corridor toward camera, framed head to knee, camera at chest height moving backward at walking pace, ceiling lights leading the eye down the frame." "Head to knee" is the key phrase — it tells the model the vertical crop before it guesses one.
3. Keep the camera still between keyframes
A single good frame is easy. Thirty matching frames is the job. The usual failure is not the characters drifting — it is the camera drifting. Keyframe 4 is shot at eye level, keyframe 5 is shot from above, keyframe 6 is a low angle, and the scene feels like three different rooms.
Fix it by writing the camera into every prompt with the same words. Pick one camera height for the scene — chest height, eye level, or shoulder height — and repeat it verbatim in every keyframe prompt. Do the same for distance: "medium shot, head to waist" in every prompt, not "medium shot" in one and "waist-up" in the next. The model treats near-synonyms as different instructions; exact repetition is what reads as continuity.
This is also where character lock cards earn their keep. A lock card fixes the face, the hair, the outfit and the defining details, so the framing prompt only has to do one job: where the camera sits and where the body sits in the frame. When one prompt tries to do both jobs at once, both slip.
4. Five framing mistakes that waste credits
- "Cinematic wide shot" in a 9:16 prompt. The model hears "wide" and squeezes a wide composition into a tall frame. Heads get cut. Never write "wide" in vertical.
- Two faces at the same height, same distance. In a narrow frame they collide. Offset one character in depth or height in every two-shot.
- Describing the room instead of the frame. "A modern office with desks, plants, glass walls and a skyline" is a location brief, not a framing instruction. Describe what the camera sees, in order, top to bottom.
- Changing the camera height mid-scene. One high angle among eye-level shots reads as a mistake, not a stylistic choice. If you want a high angle, make it a deliberate beat and return to the base height after it.
- Rendering the full scene before testing the frame. A 5-second test render of the hardest shot in the scene — usually the two-shot — tells you whether the framing language works. If the test crops a forehead, the full episode will too. Fix the words once, then render the rest.
5. From one good frame to a full episode
Good vertical framing is a system, not a talent. Same camera height, same distance words, one background element per shot, characters stacked in depth, and a test render before the full pass. That is the whole checklist, and it fits on a sticky note next to your render queue.
If you want to see the system applied to a finished scene instead of a blank page, start with the free sample: Episode 1, "Rain Check," comes with six keyframe prompts and the character lock cards for both leads, so you can run the lobby scene yourself and check the framing against the prompts line by line. The full kit extends the same method across 30 scenes — the complete urban-romance episode, with every keyframe prompt and lock card written to match.
Try the free sample — Episode 1 "Rain Check" with 6 keyframe prompts and character lock cards