Why Won't the AI Use the Camera Angle I Asked For?
Terms like "medium shot," "close-up" and "low angle" do steer image models, but they're defined against a person. On a car, a room or a prop they have nothing to measure from, so the model falls back to that object's most-photographed view. Keep the trade term, then add geometry that works on anything: how much of the frame the subject fills, where the horizon crosses it, and which edge of the frame it faces.
Why it happens
Camera vocabulary works on image models because it's all over the captions they trained on. "Medium shot," "close-up," "low angle," "over the shoulder" are real levers, so lead with them.
But each one assumes a human subject. "Eye level" means level with the subject's eyes. "Medium shot" means waist up. Point those words at a taxi and there are no eyes and no waist. When an instruction has nothing to measure from, the model gives you the most common photo of that thing, which for a car is roughly hood height, three-quarter view. We learned this over seven renders of a single taxi shot.
A few more traps:
- Distances in feet do nothing. "About fifteen feet back" is inert.
- "POV" is a genre word with no geometry in it.
- Naming a part forces the model to show it. Ask for the side of a car square to camera and the grille toward frame-right, and the car swings to three-quarter, because that's the only view where you can see the grille.
- The model doesn't flag a contradiction. It meets you halfway instead. Ask for a huge foreground curb and a close subject, and the model pushes the subject back, because the only way to make the curb big is to stand away from it.
How to fix it
- Lead with the trade term: "low angle," "wide shot."
- Give distance as frame occupancy. "The whole car fills the frame width," "her head and shoulders fill the frame."
- Give height as where the horizon crosses the subject. The horizon always sits at the camera's height. "The horizon crosses the lower edge of the doors" puts the lens low on any subject.
- Give facing as visible surfaces and a frame edge. "Side-on, passenger side to camera," or "she faces frame-left." Frame-left and frame-right are clear; "camera left" gets read two ways.
- Give depth as the plane the subject is in. Empty foreground reads as distance; a foreground object puts the subject further back.
- When it won't come together, cut words. Each new clause pulls on another lever. A shorter prompt often lands where a longer one won't.
Example
Before:
Low angle medium shot of the taxi, POV, about fifteen feet back.
After:
Low angle of a yellow taxi parked side-on, passenger side square to camera. The whole car fills the frame width. The horizon crosses the lower edge of its doors. Empty wet road in the foreground.
Every instruction in the second version means something whether the subject is a car, a door or a person.
How ShowMaker handles it
In ShowMaker's Storyboard and Shot Sheet, each shot has its own shot size and camera move, and you generate a still per shot to check the framing before any video is rendered. If a frame comes back at the wrong angle, Shoot with AI takes a plain correction ("lower the camera, make her face the camera") and redraws it. See the storyboard docs or the AI shot list generator.
Related questions
Do camera terms like close-up work in AI image prompts?
Yes, on people. Lead with them, then add geometry such as frame occupancy and horizon height so they also work on objects and places.
How do I get a low angle on an object?
Say where the horizon crosses the object. A horizon low on the subject, such as at the bottom of a car's doors, puts the camera low.
Why does adding more detail make the shot worse?
Each added clause pulls on another part of the image, and conflicting asks turn into a compromise. When a shot won't land, shorten the prompt.
Every fix on this page is a feature, not a workaround.
ShowMaker Studio is built around these problems — reference-locked casting, location look signatures, coverage-based scenes, delivery-timed captions, and per-shot model routing. Start free with 200 credits: narration over AI stills, no card.