Why your AI scene doesn't look cinematic, and six fixes

1. The light has no direction

Real footage has a light source. It comes from somewhere, it falls off, and it leaves shadow. Default AI generations light the subject evenly from the front, which reads as synthetic immediately — even to viewers who couldn't say why.

The fix is to name a source, a direction and a quality in every prompt.

Weak: cinematic lighting
Better: single warm practical lamp camera-left, low, just out of frame. Hard shadow on the right side of her face. Background falls into darkness.

2. Every shot is the same size

Twelve medium shots in a row reads as flat no matter how good each one is. Cinema varies shot size to create rhythm — wide to establish, medium to play the scene, close for the moment that matters.

Plan the mix before you generate. In a fourteen-shot scene, roughly two or three wides, a handful of mediums, several closes, and two or three inserts.

Inserts — a hand on a glass, a phone screen, a door handle — are the cheapest shots to generate and disproportionately useful in the edit. They buy you a cut point almost anywhere.

3. The camera is doing nothing, or too much

Leave camera movement unspecified and the model improvises, usually producing a slow unmotivated drift. Over-specify it and every shot swoops.

Most shots in a dialogue scene should be static. Say so explicitly — write "static camera, locked off" rather than leaving it out. Static is a decision, not an absence.

Save movement for shots where the movement means something: a reveal, a pursuit, a shift in power between two people.

4. There's no depth

Flat backgrounds with everything in focus reads as a render, not a photograph.

Specify the lens and the depth of field, and put something in the foreground:

85mm, shallow depth of field, background falls out of focus. A door frame edge in the left foreground, out of focus.

Foreground occlusion is one of the strongest signals that a camera existed in a physical space. Something was between the lens and the subject, which means the lens was somewhere.

5. Nothing is imperfect

Real footage has grain, slight underexposure, imperfect framing. AI output is clean in a way photography isn't, and that cleanliness is legible. Add grain and slight underexposure, either in the prompt or in the grade. Small amounts. The goal is to remove the sense of a perfectly rendered image, not to distress it.

6. The performance is too big

Short drama is shot close. Faces fill a vertical frame. Theatrical performance — raised voice, large gesture, broad expression — reads as overacting at that scale.

Prompt behaviour, not emotion. Naming a feeling produces a generic version of it.

Weak: she is furious
Better: jaw tight, does not raise her voice, holds eye contact, does not blink

What order should I check these in?

When a scene isn't working, diagnose in this order — highest impact first.

Does the light have a direction? Is there shot-size variety? Is the camera deliberately static, or deliberately moving? Is there foreground and depth? Is it too clean? Is the performance too big?

Most scenes are fixed by the first two.