The difference between AI footage that looks like a slot-machine pull and footage that looks like your movie is control. Everything below is about taking control — so you burn fewer generations, get more keepers, and spend less doing it. Failed generations never charge you, but a keeper on the first pull is cheaper than a keeper on the fourth.
Write the shot like you’re briefing your DP
The prompt is a shot brief, not a wish. Vague in, vague out. Name four things every time:
- Subject — who or what, specifically. “A man” is weak. “A weathered dockworker in his 50s” is a shot.
- Action — one clear action per generation. One shot, one beat. Don’t ask for a whole scene.
- Camera — angle, lens feel, movement. “Slow push-in, low angle.” “Handheld, tracking behind.”
- Light — time of day, source, mood. “Hard side light, late afternoon.” “Sodium streetlight, wet pavement.”
Short and specific beats long and flowery. You’re directing, not describing a painting.
Use references — this is the biggest lever you have
Prompt-only generation asks the model to imagine everything. References tell it what you want. Upload reference images for faces, style, or composition, and the model anchors to them instead of guessing.
Name your references. Instead of leaving them as @image1, give them useful names — @Sarah, @warehouse, @the-knife — and call them out in the shot brief: “@Sarah steps into @warehouse, backlit.” Numbered order still works as a fallback, but naming keeps a complex shot straight.
Faces are special. Any real person’s face has to belong to a performer you’ve cleared, with a signed release on file. Once cleared, that face can be reused within that production. Name it after the performer and it becomes self-documenting.
Match motion with a video reference
Some shots — a fight, a fall, a specific gesture — are hard to direct with words alone. Feed the Generate room a clip of the movement you want, shot rough on a phone or sourced from material you have the right to use. The movement becomes an anchor instead of a guess.
Iterate with takes — on the real thing
AI generation produces variations. Use takes to make two to four passes from the same brief, then pick the keeper.
Don’t draft on a cheaper model and expect a premium model to reproduce it. Different models produce different shots, not higher-resolution versions of the same shot. Iterate at the resolution you intend to deliver.
Pick resolution and length for the job
- 720p — proofing, web, and previz.
- 1080p — the production standard for most finishing.
- 4K — premium finishing for hero shots.
Generate a little long and trim in the edit. Greenlit Dark offers 5, 10, and 15-second shots: choose enough room for the action to land without paying for footage you know you won’t use.
Clear the floor, or it doesn’t ship
- Real faces and voices need signed releases. No release, no likeness.
- No third-party IP. Don’t submit characters, franchises, or settings you don’t control.
- Hard content still needs your paper. Gore, violence, and practical-effect extensions are reviewed in the context of your approved production.
Every delivered shot passes human review before the master unlocks. That review is part of what makes the footage usable in a real production.
The short version
Be specific. Use and name your references. Match motion with a clip. Take a few passes and keep the best. Clear your paper up front. Those five habits improve your keeper rate, control spend, and make the result look like your movie instead of a machine’s guess.
Questions on a specific shot? Email support@greenlitdark.com — we make movies too.