How to write a prompt that works
The same idea gets a different result on every model. This is why, and what to write instead.
2 min read · last checked
The model is not the problem, the prompt usually is
People compare two platforms, get a better result on one of them, and conclude the other has worse models. Most of the time both run the same model. What differs is the sentence that went in.
These models were trained on text written by people who describe pictures for a living: photographers, cinematographers, art directors. They respond to that language because they have seen a lot of it. "Blurry background" is not how anyone in that trade writes. "Shot at f/1.8, shallow depth of field, the background falling well out of focus" is, and it lands differently.
Write a brief, not a list of words
A row of comma-separated keywords was how the first generation of image models worked. The current ones read whole sentences and weigh the order. A written-out description of three or four sentences beats fifteen keywords almost every time.
Say it in this order, because the beginning weighs heaviest: the subject and what it is doing, the framing and the lens, the light and where it comes from, then the colour treatment and the material detail that makes it read as real.
One thing that genuinely does not help: stacking adjectives. Several model makers say so in their own documentation. Six style words fight each other, and the model resolves that fight by ignoring most of them.
Every family wants it a bit differently
This is the part that is hard to know without testing, so the studio carries it for you. Each model in the catalogue has its prompt style, its strengths and an example prompt written for it, with the source it came from. The assistant reads the same notes, and so does the prompt enhancer.
Roughly, the families split three ways. Some read your prompt as a series of directions and get better the more you tell them. Some want motion first and treat the rest as background. And some want you to keep it short, because past a certain length they start listening less rather than more.
Let the studio write the technical half
The camera moves, lens settings, light and colour are the part that reads as craft, and they are also the part that is the same every time. The studio has 32 of them as presets: you click one, and the exact wording is added to your prompt.
They are added word for word, not paraphrased. That is the whole point of a preset: the phrasing is the thing that works, so it has to survive intact all the way to the model.