Techniques
How to write AI image prompts
Good image prompts name the subject, the setting, the lighting, the medium and the framing, in roughly that order, using concrete nouns rather than praise.
The most useful thing to know about prompt writing is that adjectives of quality do almost nothing. Words like "beautiful", "amazing", "masterpiece" and "8k" are so evenly distributed across training data that they carry very little direction. What carries direction is specificity: a named lighting setup, a named medium, a named lens, a named palette. "Soft window light from the left" changes the image; "good lighting" does not.
A structure that reliably works is subject, then action or pose, then setting, then lighting, then medium and style, then framing. Early tokens carry more weight in most models, so lead with what the picture is actually of. "An elderly fisherman mending a net on a harbour wall at dawn, soft directional light, oil painting with visible brushwork, chest up" gives the model six separate hooks, each of which does real work.
Framing and composition are the most under-used part of a prompt. Saying "chest up", "wide establishing shot", "low angle", "centered", or "lots of negative space on the right" gives you control over something that otherwise gets decided randomly, and it is usually the difference between an image you can use and one you cannot. If the picture has a job, whether that is holding a title or fitting a profile circle, say so.
Finally, iterate one variable at a time. Lock the seed, change only the lighting clause, and see what happened. If you rewrite the whole prompt between generations, and the seed changes too, you learn nothing about which change mattered. Most of the skill here is in running controlled experiments rather than in knowing magic words.
In practice
- Weak
- a beautiful portrait of a fisherman, masterpiece, 8k, highly detailed
- Strong
- an elderly fisherman mending a net on a harbour wall at dawn, soft directional light from the left, oil painting with visible brushwork, muted earth palette, chest up
- Why
- Six specific hooks instead of four pieces of praise
Worth remembering
- Concrete nouns beat quality adjectives; drop "masterpiece" and "8k"
- Order matters: subject, action, setting, lighting, medium, framing
- Name the framing explicitly, especially if the image has a job to do
- Lock the seed and change one clause at a time when refining
Try it in the generator
The advanced panel exposes the negative prompt, scheduler, inference steps and seed. Twenty images free, no account needed.
Start generatingFrequently asked questions
- How long should a prompt be?
- Long enough to name subject, setting, lighting, medium and framing, which is usually 20 to 40 words. Beyond that, later terms get progressively less weight, so extra length has diminishing returns.
- Do words like 8k and masterpiece help?
- Very little. They appear across such a wide range of training images that they barely steer anything. Replace them with a specific medium, lens or lighting description.
- Should I name a famous artist?
- Avoid naming living artists; it raises real ethical objections and is blocked in many tools. Describing the qualities you want, such as "loose ink linework with watercolour washes", gets you the look without imitating a person.
Related terms
What is a negative prompt?
A negative prompt is a second description of what you do not want, which the model actively steers away from during generation.
What is a seed in AI image generation?
The seed is the number that determines the random starting noise, so the same seed and prompt reproduce the same image every time.
What is CFG scale (guidance scale)?
CFG scale, or guidance scale, controls how strongly the model is pushed toward your prompt and away from the negative prompt at each step.
Back to the glossary, or apply it on a use case or an art style.