Prompting for cinematic video: 7 tricks stolen from DPs
Real cinematographers don't say 'hyperrealistic 8K unreal engine'. Here are seven concrete prompting techniques we've lifted from DPs — and the before/after they make on your output.
The best AI prompts look almost nothing like the ones trending on social. Forget "hyperrealistic, 8K, unreal engine, intricate details, masterpiece". Real cinematographers don't talk like that. And — it turns out — modern video models don't listen to that either.
We spent a week with a working DP, rewrote 40 prompts using her vocabulary, and re-ran every one through our pipeline. Every single output was better. Not *differently weighted* — unambiguously better. Here's what she told us.
1. Lead with the camera, not the subject
"Handheld, 35mm, slow push-in on —" is better than "a man sitting —". Give the model a lens, a motion, and a framing *first*. The subject comes after, because the camera is what the model is actually simulating.
2. Name the light source
"Golden-hour sun from frame right" or "soft key, hard backlight" gives the model a real physical scene to reason about. Light isn't decoration — it's half the image.
If you only add one thing to your existing prompts this week, add the light. You will see the difference on clip #1.
3. Specify the crop
Close-up, medium, wide, aerial. The model picks the wrong one surprisingly often if you don't say.
4. Describe the air
Dust motes. Haze. Rain. Snow. Fog at ankle-height. Smoke rolling over a shoulder. These aren't decoration — they *tell the engine where to put depth*. A "cinematic" shot without atmospheric stuff in it is a flat shot. The air between the subject and the camera is what makes it feel like film.
“If you can see the air, you can feel the room.— An old DP quote, still the best one
5. Use one reference word, not a list
"Fincher-dark" beats "moody, gritty, cinematic, atmospheric, dramatic". Lists of adjectives flatten each other — the model averages them and you end up with the blandest thing that satisfies all five.
Pick the one word that carries the visual culture, and let it do the work.
6. Skip the "masterpiece, trending, 8K" junk
They don't help modern video models. They used to help older image models on Civitai three years ago. Different era. Including them now is noise — every token you spend on junk is a token not spent on something useful.
This is the single most common mistake we see in prompts pasted from Discord. Copy-pasting 2022-era image prompts into 2026 video models genuinely makes them worse, not better.
7. End with what the shot *isn't*
"No text, no watermark, no people" is faster than trying to coax them out in post. Negatives are cheap, post is expensive.
Putting it together
A full "DP-style" prompt looks like this:
> *Handheld, 35mm, slow push-in on a woman tying her running shoes on a wet concrete step. Pre-dawn blue hour, practical streetlamp key from frame left, soft fog between camera and subject. Medium close-up. Fincher-dark. No text, no watermark.*
One lens, one motion, one light source, one atmosphere word, one reference, one framing, one exclusion list. That's the whole technique.
Paste one of your old prompts, rewrite it in DP form, run both on your trial tokens. Keep whichever you'd actually ship.