Writing Image Prompts for Midjourney, DALL·E & Stable Diffusion

Type "a cat" into any image generator and you'll get a cat. It might be a photo, a cartoon, an oil painting, or something that looks like clip art from 2004. The model guessed, because you gave it almost nothing to work with. The gap between a forgettable result and one you'd actually use comes down to how much you tell the model and the order you tell it in. This guide walks through what goes into a prompt, then shows how the three big tools each want that information delivered.

The parts you build a prompt from

Think of a prompt as a stack of decisions. Each one narrows the space of images the model could produce. Leave a layer out and the model fills it in for you, usually with something generic.

A reliable order to work through:

• Subject — who or what is in the frame, and what they're doing. "A fox curled asleep" beats "a fox."
• Style or medium — photograph, watercolour, 3D render, charcoal sketch, cyberpunk illustration. This decision colours everything else.
• Composition — close-up, wide shot, bird's-eye view, subject placed off-centre, foreground and background elements.
• Lighting — golden hour, soft window light, harsh noon sun, neon glow, rim lighting. Lighting does more for mood than almost anything else.
• Mood and colour — melancholic, playful, tense; a warm amber palette, muted pastels, high-contrast monochrome.
• Camera and lens — for photographic looks: 35mm, 85mm portrait lens, shallow depth of field, macro. These push the model toward realism.
• Detail and quality modifiers — "intricate texture," "sharp focus," "highly detailed," and tool-specific quality flags.

You don't need every layer every time. A children's-book illustration rarely needs a lens spec. But moving top to bottom keeps you from burying the subject under adjectives, which is one of the most common ways prompts go sideways. If you want the broader thinking behind structuring instructions for any model, prompt engineering fundamentals covers the same logic applied beyond images.

Midjourney: parameters carry the weight

Midjourney reads a comma-separated description and then takes instructions from parameters tacked onto the end. The description sets the scene; the flags control the machinery.

The ones you'll reach for constantly:

• --ar sets aspect ratio (--ar 16:9 for widescreen, --ar 2:3 for a portrait poster).
• --v picks the model version.
• --style raw dials back Midjourney's tendency to prettify and gives you a more literal read of your words.
• --no removes elements: --no text, watermark.
• :: weights terms. forest::2 fog::1 tells the model the forest matters twice as much as the fog.

Weighting and --no are your steering wheel. If neon signs keep bleeding into a scene you wanted clean, --no neon clears them faster than adding ten "no neon" words to the description.

DALL·E 3: write it like a sentence

DALL·E 3 was built to read plain, descriptive language, so keyword soup works against you here. Write full sentences the way you'd brief a human illustrator, and it will handle the composition intelligently. It also rewrites your prompt behind the scenes to add detail, which is helpful until it isn't — so be explicit about anything you feel strongly about.

Because DALL·E leans on grammar and relationships between things, spatial phrasing lands well: "a red mug to the left of an open notebook, a pair of glasses resting on the pages." Describe the arrangement and it usually respects it. Save the technical shorthand for the other two tools. If you're coming from writing text prompts, many of the same habits transfer — see how to write AI prompts for the general approach.

Stable Diffusion: keywords, weights, and negatives

Stable Diffusion works closest to the bare machinery. Prompts are comma-separated keywords, and you get direct control over emphasis and exclusion. This is the tool where a negative prompt earns its keep, because a well-populated one quietly fixes the recurring flaws these models produce.

Emphasis uses parentheses and numeric weights. (dramatic lighting:1.3) boosts a term; (background clutter:0.6) suppresses it. Stack a few and you're sculpting.

A starter negative prompt to keep on hand:

• Anatomy fixes: extra fingers, fused hands, deformed limbs
• Junk artefacts: watermark, signature, text, jpeg artifacts
• Style corrections: cartoon (when you want a photo), oversaturated, blurry

Reuse and tune it per project rather than rewriting from scratch. When you want to assemble one from a checklist, a negative prompt builder saves the guesswork, and a general image prompt helper can lay out the positive side in the right order.

Aspect ratio, seeds, and staying in one language

Aspect ratio isn't cosmetic. A square crop composes differently from a tall portrait, and the model frames the subject to fit. Decide the shape before you polish the details, because a wide 16:9 and a tall 9:16 of "the same" prompt often want different compositions entirely.

Seeds are how you stop fighting randomness. Every generation starts from a seed number; reuse it and the model begins from the same point, so your next tweak changes only what you edited instead of rerolling the whole image. The workflow that actually works:

• Generate a batch, find one that's 80% there.
• Grab its seed.
• Lock the seed and adjust a single element — the lighting, one object, the palette.
• Compare against the original.

That turns "spin the wheel and hope" into something closer to editing.

On language: the models were trained overwhelmingly on English text paired with images, so English descriptions map more precisely onto what the model actually learned. Prompts in other languages often work, but you lose fidelity on nuanced or uncommon terms, and niche style words may not register at all. If your first language isn't English, write the idea in your own words, then translate the final prompt before generating.

Putting it together

Two habits separate people who get consistent results from people who keep rerolling. First, build every prompt in the same order so you never forget the lighting or leave the composition to chance. Second, treat each generation as a test of one idea, not a lottery ticket.

A worked example, adaptable across tools:

Read that back against the layers: subject (dewdrop on a web), style (macro photograph), composition (single droplet, blurred background), lighting (backlit sunrise), colour (warm gold against dark green), detail modifiers, and the aspect ratio at the end. Nothing is left to a guess.

Most weak results trace back to a missing layer or a subject smothered in adjectives, and both are easy to catch once you know the pattern — common prompting mistakes collects the ones worth avoiding. Get the anatomy right, respect what each tool wants, and change one thing at a time.

References

• Midjourney
• Stability AI