A prompt is the description you type. It is the biggest lever you have: a better prompt improves results far more than any setting does.
The good news is that it is ordinary writing. There is no syntax to learn and no magic words.
Start with the shape that works
Most good prompts have four parts, in roughly this order:
Subject — setting — light — treatment.
a red fox asleep on a stack of old library books, in a dusty study,
warm afternoon light through a window, shallow depth of fieldYou do not need all four every time. But when a picture comes out flat or generic, the missing part is almost always light or treatment.
| Part | Answers | Examples |
|---|---|---|
| Subject | what is in it, doing what | a fox asleep, a woman laughing, a ceramic mug |
| Setting | where it is | a dusty study, an empty beach, a white background |
| Light | how it is lit | warm afternoon light, harsh flash, candlelight, overcast |
| Treatment | what kind of picture it is | oil painting, film photograph, pencil sketch, 3D render |
Be specific about what you care about
Every word you leave out is a decision made for you.
| Vague | Specific |
|---|---|
| a dog | a wet golden retriever shaking off water |
| a nice kitchen | a narrow galley kitchen, morning sun, brass taps |
| professional photo | studio product photo, white seamless background, soft box lighting |
Be specific about the things that matter to you and leave the rest open. A prompt that specifies everything down to the doorknobs gives you less room to be pleasantly surprised, and models handle long lists of constraints poorly.
Say what you want, not what you do not
“A street with no cars” often produces a street with cars, because the description still contains the word cars. Describe the thing you do want — “an empty street at dawn” — instead.
If there is a negative prompt box, that is the correct place for things to avoid. Putting them in the main prompt tends to summon them.
Length
Two or three lines is the sweet spot. Long enough to say subject, setting, light and treatment; short enough that nothing gets lost.
Very long prompts are not more powerful. Past a certain point, later words get less attention than earlier ones, so a hundred-word prompt often ignores its own ending. Put what matters most first.
Change one thing at a time
The single most useful habit. Generate, change one word, generate again.
If you change the subject, the lighting and the style at once and the result is worse, you have learned nothing about which change did it. Change one word and you learn something you can use every time after.
Reuse what worked
Recently used prompts appear as chips near the prompt box — click one to bring it back rather than retyping it.
When something comes out well, keep the prompt. A prompt that reliably produces the look you want is worth more than any individual picture, because you can apply it to a hundred different subjects.
Things models are bad at
Worth knowing before you spend credits fighting them:
- Text inside the image. Signs, labels and logos come out as convincing-looking nonsense. Some models are better than others; none are reliable. Add real text afterwards in any editor.
- Exact counts. “Seven birds” gives you some birds. Small numbers work better than large ones.
- Hands and teeth. Improving, still awkward. A hand holding something simple works better than one on its own.
- A specific real person’s face. Not reliable, and often restricted.
- Precise spatial arrangements. “The cup exactly to the left of the book, behind the lamp” is usually approximated rather than followed.
A worked example
Starting point, and what each round adds:
a catA cat. Could be anything.
a ginger cat sitting on a windowsillNow there is a scene.
a ginger cat sitting on a windowsill, rain on the glass, grey afternoon lightNow there is a mood.
a ginger cat sitting on a windowsill, rain on the glass, grey afternoon light,
35mm film photograph, shallow depth of fieldNow it looks like a photograph someone took on purpose.
Four rounds, one addition each. That is the whole method.