How to write scene prompts for reference video
When a video model already has reference images of your character or product, the prompt has a different job. It describes the scene, not the subject. A good scene prompt names the setting, one main action, the camera move, the light and the ending, in that order, in two to four plain sentences. This guide gives you that structure, worked examples, and the habits that quietly break consistency.
Why reference prompts are different
In text-to-video, the prompt carries everything, so people learn to pack it with appearance details. With references, those details compete with the images. If the references show short brown hair and the prompt says auburn waves, the model averages the two and you get a third version. The rule is simple: identity lives in the references, everything else lives in the prompt.
That frees the prompt for what references cannot provide: where the shot happens, what moves and how it is filmed.
The five-part structure
1. Setting. One sentence on place and time: "A small bakery kitchen at dawn, flour on the steel counter." Specific nouns beat adjectives. "Steel counter" tells the model more than "modern kitchen".
2. Action. One main action per clip, in present tense: "She lifts a tray of bread from the oven and sets it on the counter." Two actions in sequence are fine for clips of eight seconds or more. Three or more usually produces rushed or skipped movement.
3. Camera. Name the shot size and the move: "Medium shot, slow push-in." "Close-up, handheld, slight sway." "Wide shot, static." If you leave the camera out, the model picks one, and it picks differently each time, which makes clips hard to cut together.
4. Light. Name the source and direction: "Warm light from a window on the left." Light is one of the main triggers of face drift, so describe it plainly and keep it similar across clips that will be edited together.
5. Ending. Say where the clip ends: "Ends as she looks up at the camera and smiles." A defined end frame makes it easier to cut to the next shot.
Worked examples
Character, everyday shot: "Stands in a minimal kitchen in warm morning sunlight, pours coffee from a glass carafe into a white mug, then looks up and smiles at the camera. Medium shot, soft natural light." This is close to the prompt behind the clip on this page.
Product, hero shot: "The bottle stands on wet black stone, small drops on the glass. A slow orbit around it, low angle. A single hard light from behind draws a rim along the edges. Ends on the front label facing the camera."
Character, talking shot: "A home office with bookshelves behind. He leans toward the camera and explains something with one hand. Medium close-up, static camera, eye level. Soft daylight from the right. Ends on a short pause and a nod."
Notice what is missing from all three: no hair color, no eye color, no clothing, no logo description. The references handle those.
Length, shots and format
Length changes what a prompt can hold. Kling 3.0 Omni generates up to 15 seconds per request and supports multi-shot generation, according to Kling's guide as of October 2026. Veo 3.1 reference clips are always 8 seconds, per Google's documentation. For a three- or five-second clip, write one action. For ten seconds or more, write two, or use separate shot descriptions if the model supports multi-shot.
When two subjects share a scene, refer to each by its tag, such as @Element1 and @Element2 in Kling's Omni models, so the model knows who does what. Aspect ratio is a setting, not prompt text. Choose 9:16 for vertical social video and 16:9 for web and YouTube rather than writing "vertical" into the prompt.
Common mistakes and checklist
Repeating appearance details from the references. Style words that fight the references, such as "anime" when the references are photos; if you want a stylized look, make stylized references. Quality padding like "masterpiece, 8k, ultra detailed", which makes prompts harder to compare. Changing several things at once when a clip fails; change one part per attempt so you learn what caused the problem.
Checklist: setting in one sentence; one action per five seconds; shot size and camera move named; light source and direction named; an end frame; no appearance details; aspect ratio set in settings.
RefVid is set up for this split. The character or product is saved once with its references, and the prompt box, up to 2,000 characters, is for the scene only. You do not need to tag the subject; RefVid attaches the asset to the prompt for you on Kling 3.0 Omni.
Frequently asked questions
- How long should a scene prompt be?
- Two to four sentences, roughly 30 to 70 words. Long enough to cover setting, action, camera, light and ending; short enough that every word does a job.
- Should I describe my character's appearance in the prompt?
- No. When reference images already show the character, appearance words compete with them and cause drift. Describe the scene instead.
- Can I use negative prompts with reference video?
- Some models and APIs accept a negative prompt field. It can suppress specific artifacts, such as on-screen text, but it does not fix identity drift. Better references do.
- What does it cost to try a prompt on RefVid?
- 42 credits per second: 126 for a 3-second draft, 210 for 5 seconds and 420 for 10. The first 50 accounts get 630 free credits. Later accounts get 126.