How to write image prompts

Why are image prompts written differently from chat prompts?

Because the models work differently. A chat model reads your prompt as language and follows instructions. A diffusion model does not follow instructions at all — it steers an image toward a set of weighted concepts. Grammar carries almost no information for it; the individual descriptive terms carry nearly all of it.

The practical consequence is that a well-written paragraph performs worse than a comma-separated list of the same concepts. Words like "with", "featuring" and "in the background" mostly consume weight without contributing.

Weaker:  A beautiful photograph of an old fisherman who is mending
         his nets while sitting on a harbour wall at sunset.

Stronger: elderly fisherman mending nets, stone harbour wall, golden
         hour, shallow depth of field, 35mm film, muted palette,
         three-quarter portrait, high detail --ar 3:2

What order should an image prompt be in?

Terms nearer the front of a descriptor stack influence the result more, so the order is not cosmetic. The convention that every major model’s documentation converges on:

1. Subject — what the image is of. The single most important slot.
2. Action — what the subject is doing.
3. Environment — where it is happening.
4. Composition — framing, angle, depth of field.
5. Lighting — golden hour, studio softbox, harsh noon, rim light.
6. Medium and style — 35mm film, oil painting, isometric 3D render.
7. Quality modifiers — sharp focus, high detail.
8. Parameters — aspect ratio, seed, model-specific flags.

If the subject is coming out wrong, the fix is almost always to move subject detail earlier and delete adjectives from the tail, not to add more words.

What should go in a negative prompt?

A negative prompt is a separate input listing what to exclude. Crucially, it is separate — writing "no text, no watermark" inside the main prompt often has the opposite effect, because the model weights the words it sees regardless of the word "no" in front of them.

The reliably useful negatives are artefacts rather than content: text, watermark, signature, extra fingers, deformed hands, blurry, jpeg artifacts, oversaturated. Excluding content you simply did not mention is usually unnecessary.

How do you get a consistent style across several images?

• Fix the seed. A seed makes generation reproducible, so a prompt change becomes the only variable. Without it you cannot tell whether an edit helped or you simply got a different roll.
• Reuse the tail verbatim. Keep the lighting, medium and quality section identical across a set and vary only the subject.
• Use a style reference image where the model supports it. It communicates palette, texture and lighting far more precisely than adjectives.
• Describe rather than name living artists. "Loose ink linework with flat washes" is both more precise and less ethically fraught than naming someone whose living depends on that style.

How do you describe lighting and composition usefully?

Borrow the vocabulary of photography and film, because that is what the training data is labelled with. Specific terms outperform impressionistic ones by a wide margin.

Note that "cinematic" and "detailed" are the image-prompt equivalents of vague qualifiers — they feel like specifications and steer almost nothing.

What does a complete image prompt look like?

Prompt:
product shot of a matte black ceramic mug, steam rising, on a
weathered oak table, shallow depth of field, single softbox from
camera left, dark neutral background, 85mm, high detail, sharp
focus --ar 4:5

Negative prompt:
text, watermark, logo, hands, oversaturated, plastic sheen

Every term earns its place: subject and material, action, environment, composition, a single named light source, background, lens, quality, ratio. The negative prompt covers artefacts, not content.