Camera language AI understands, and the terms that quietly do nothing

Write a prompt full of proper film terminology and it reads authoritative. Whether any of it reached the model is a separate question, and the honest answer is that a substantial portion did not.

These models learned from captioned footage. A term only works if it appeared in those captions attached to footage that actually demonstrated it. Terms that circulate among filmmakers but rarely appear in a caption have no handle at all, no matter how precise they are as language.

So the working vocabulary is not the vocabulary you learned on set. It overlaps with it, and the gaps are worth knowing rather than discovering one failed generation at a time.

The reliable core, across essentially every model:

• Static locked-off shot — no movement. Explicitly requesting this is the strongest camera instruction available, and the most underused.
• Slow push in / slow pull out — translation along the lens axis.
• Pan left / pan right — horizontal rotation.
• Tilt up / tilt down — vertical rotation.
• Tracking shot alongside — lateral translation following a subject.
• Handheld — adds organic instability. A texture, not a move.

Each of these describes a physical thing a camera body does, in a direction, at a speed. Add the speed — 'slow' does real work, and unqualified moves consistently come out faster than intended.

Compound moves. A dolly zoom requires the camera to move while the lens changes focal length in the opposite direction. Two coordinated changes; models produce neither cleanly. Same for a move that changes direction partway.

Equipment names. Crane, jib, dolly, Steadicam, gimbal. These describe what the camera is mounted on, not what it does. A crane shot might mean rising, descending, or an aerial view depending on what the model saw in training. Say the movement instead: 'the camera rises slowly'.

Abstractions. 'Cinematic camera movement', 'dynamic camera work', 'epic sweeping shot'. These are mood words dressed as technical terms. They occupy prompt space and displace instructions that would have worked.

Stating a focal length genuinely changes the image. A 24mm gives you the wide framing, the perspective stretch and the close-subject distortion you would expect; an 85mm gives compression and a tighter frame. This works because focal length is one of the few technical values that gets written into photo captions routinely.

Aperture is weaker. Writing f/1.4 sometimes increases background blur and often does nothing measurable. If you want shallow depth of field, ask for it in plain language — 'the background falls out of focus' — rather than in stops.

Also effective, because they appear in captions: anamorphic (changes the flare and the bokeh shape), named film stocks (real handles on grain and colour), and macro (changes subject distance decisively).

Largely ineffective: shutter angle, ISO, specific camera bodies, sensor sizes. These are production metadata that rarely made it into a description of what the picture looks like.

Light is where these models are strongest, probably because describing light is what captions actually do. Almost everything works, provided you say it plainly:

• Direction — 'lit from the left', 'backlit', 'lit from below', 'overhead'.
• Quality — 'hard', 'soft', 'diffused', 'a single source'.
• Named conditions — 'late afternoon sun', 'overcast', 'fluorescent', 'candlelight'. These carry colour temperature, contrast and direction together, and they are efficient.

The one instruction worth adding to almost every prompt: a single key light. Models default towards flat, evenly lit, multi-source images. Specifying one source is what produces shape and shadow, and it costs four words.

Take one prompt. Generate it with no camera instruction. Then generate it again changing only the camera term, one term at a time, keeping seed and everything else fixed.

Twenty generations gives you a real answer about what your current model responds to — and it is genuinely different between models, which is why borrowed lists get you started and no further.

Keep the results. When a new model arrives, re-run the same test rather than assuming the vocabulary transferred. The categories in this guide hold up well across releases; the specific terms shift more than you would expect.