Mistakes in image prompts and how to fix them

Ten common mistakes in prompts for image generation, three before-and-after examples and a checklist to run before every generation.

Image prompt engineering Updated

In short

Most bad results from image generators come from a few repeating mistakes in the prompt, not from the generator. The prompt describes only the subject and leaves style, light and framing to chance; it is padded with words like “masterpiece” and “8k” that do almost nothing; it contradicts itself; it tells a story instead of describing one moment; it asks for long text inside the picture; or it changes everything at once between attempts. Each of these has a simple fix: describe a shot, choose one option for each part, keep text for the layout and change one thing at a time.

10 mistakes in image prompts

Each with what it leads to and how to fix it.

MistakeWhat happensFix
Only the subject a random style and light describe the whole shot
Words about quality almost nothing changes concrete visual words
Contradictions a muddled compromise one option for each part
A story instead of a moment a strange collage of events one moment that tells it
Long text in the picture broken letters text in the layout
“No people” at the end people appear anyway describe what should be there instead
Too many objects some are lost or merged three or four main objects
No format cropping ruins the composition the format in the prompt
Names of living artists legal and ethical risk describe the style in words
Changing everything at once nothing is learned one change per attempt

Before and after: 3 fixed prompts

Three typical mistakes and how the same idea looks as a good prompt.

Only the subject and “quality”

Words about quality are replaced by framing, light and colours.

interior.txt
# Before: only the subject and words about quality
beautiful modern living room, high quality, 8k, masterpiece

# After: a shot instead of a wish
A modern living room with a light grey sofa and an oak coffee table,
interior photography, wide shot at eye level, 24mm lens,
overcast daylight from large windows, white walls and warm wood tones,
calm and airy mood, 16:9 format

Contradictions

Every pair of opposite words is replaced by one decision.

poster.txt
# Before: contradictions — the generator chooses at random
minimalist poster with lots of details, night scene in bright sunlight,
realistic cartoon style

# After: one decision for each part
Minimalist poster of a lighthouse at night, a single beam of light,
flat graphic illustration with three colours: navy, cream and orange,
large empty sky in the upper half, centred composition, 2:3 format

A story instead of a moment

Three events become one frame with a detail that hints at the rest.

moment.txt
# Before: a story — a picture cannot show what happens next
a man comes home after a long day, feeds his cat,
then remembers it is his birthday and smiles

# After: one moment that tells the story
A tired man in a coat kneels in a warm hallway and smiles
at a ginger cat rubbing against his hand, a small birthday cake
on the shelf behind him, documentary photography, medium shot,
warm lamp light, cosy evening mood, 3:2 format

A checklist before generating

  1. 01

    Is the subject named concretely?

    “An oak dining table” instead of “furniture”.

  2. 02

    Are style and medium chosen?

    Photograph, illustration, 3D — one of them.

  3. 03

    Are framing and angle set?

    Close-up or wide shot, from which height.

  4. 04

    Is the light described?

    Its type and direction, not “good lighting”.

  5. 05

    Are there no contradictions?

    Read the prompt once more looking only for opposites.

  6. 06

    Is it one moment?

    If there is “then” in the prompt — it is a story.

  7. 07

    Is the format set?

    Square, vertical or wide — for the place the picture will go.

  8. 08

    Is there room for text?

    If the picture goes on a banner — the empty side is named.

Questions about mistakes in image prompts

Why does the generator ignore part of my prompt?

The prompt is too long or contradicts itself; the end gets the least attention — the main thing goes first.

Why are hands and text broken?

These are the hardest places for generators; simpler poses, short words and several attempts help.

Do negative words work?

Often badly: “no people” can bring people in. It is better to describe what should be in the frame.

Do “8k” and “masterpiece” help?

In modern generators — almost not; they take the place of words that matter.

How many attempts are normal?

Several per idea: the first ones show what the generator understood, the next ones refine it.