A prompt is not a description. It's a budget. Every word you spend on one thing is a word not spent on another, and the model has a fixed amount of attention to divide.
Most bad couple-video prompts aren't badly written. They're over-allocated — a paragraph about the mood and four words about who is in the frame.
The budget model
Think of attention as roughly fixed. If you describe fog, cobblestones, a flickering lamp, a torn curtain, and dust motes, most of that attention went to the room. The faces got the remainder.
The room doesn't matter. The faces are the entire product.
So the first rule of prompt weighting isn't syntax. It's subtraction: decide what the clip is about, and spend almost everything there.
The one thing that must win
For a two-person zombie clip, that one thing is: two specific people, together, in one readable moment.
Not the apocalypse. Not the mood. Two faces that stay themselves while something happens to them.
If you had to compress a prompt to one sentence, it would be something like: "Two people turn toward each other in a dim room, both clearly the same two people from the photos."
Everything you add after that is decoration that competes with it.
What to cut first
In order:
- Camera language. "Cinematic, 35mm, shallow depth of field, anamorphic" — these describe a look the model partly applies anyway, and they eat attention.
- Adjectives about atmosphere. "Eerie, haunting, unsettling." Mood words are the least actionable tokens in any prompt.
- Background detail beyond one anchor. One object sets the room. Five objects fight each other.
- Anything describing the source photos. The model already has them. Repeating "a man with short hair and a woman with long hair" is paying twice for information you gave it for free.
What survives: who, doing what, where, in what light.
Order matters more than emphasis
Models weight early tokens more heavily than late ones. This is a coarse, well-documented bias, and it's more reliable than any emphasis syntax.
So put the subject first. "Two people turn toward each other" beats "In a fog-filled ruined street at dusk, two people turn toward each other" — the second one starts with the street.
If you have one sentence of budget, spend it on the subject and the action. Location and light go second. Decoration goes last, or nowhere.
About weighting syntax
Brackets, colons, numeric weights — (term:1.3), [term], ++term — are conventions of specific tools. A number that means "more" in one model is literal text in another, and in some pipelines it's simply stripped.
Two practical consequences:
- Don't copy weighting syntax from a prompt written for a different product.
- Don't rely on weights to rescue a prompt whose priorities are wrong. If the subject is buried in paragraph three, no colon fixes that.
Rewriting for priority beats weighting almost every time.
A worked example
Start with what people usually write:
A haunting, cinematic scene in a fog-covered abandoned city at dusk, ruined buildings, flickering streetlights, dust in the air, anamorphic lens, shallow depth of field, two people walking slowly, eerie atmosphere, unsettling mood.
Now the same intent, re-allocated:
Two people walk slowly toward the camera through fog, holding hands, both faces clearly lit and visible. Abandoned street behind them.
Same scene. Half the words. The second one will hold identity better, because identity is what it spends on.
How to tell it worked
Compare two renders and ask a single question: did the faces come back as the same two people?
If yes, the budget was right. If the scene is beautiful and the faces drifted, you spent on the wrong thing — and no amount of prompting will fix a prioritisation error.
Try it at the generator; price at pricing.
Frequently asked questions
Should I write a long prompt or a short one?
Short, unless the extra length is all about the two people and what they do. Length itself isn't the problem; misallocation is.
Does putting a word first make it stronger?
Generally yes. Early tokens carry more weight than late ones across most of these models.
Do numeric weights work?
In some tools, in some versions. They're not portable, and they're a poor substitute for putting the important thing first.
Can I describe the people from the photos?
You can, but it's redundant — the images are the reference. Use the words for action and staging instead.
Why did my detailed prompt give a worse result?
Because detail is attention spent. A prompt that describes everything equally describes nothing in particular.