A Midjourney prompt is a compact brief. The model reads your words roughly left to right and weights the earlier ones more heavily, so the order and specificity of what you write changes the image as much as the subject itself. This guide breaks a prompt into its parts, explains every parameter you are likely to reach for, and shows finished examples you can adapt.
The anatomy of an image prompt
Most strong prompts move through the same six or seven ideas, in this order. You do not need all of them every time, but naming each one gives the model fewer decisions to make on your behalf.
- Subject - who or what the picture is of. Be concrete: a snow leopard beats an animal.
- Descriptive detail - age, clothing, colour, texture, number, action. "a snow leopard mid-leap, thick winter coat, pale grey with black rosettes".
- Style or medium - photograph, oil painting, watercolour, 3D render, line art, pixel art, cyanotype. This is the single biggest lever on how the image feels.
- Lighting - golden hour, overcast, hard studio flash, rim light, neon, candlelit. Light does most of the mood work.
- Composition and framing - wide shot, close-up macro, top-down, low angle, rule of thirds, centred, shallow depth of field.
- Mood and colour - moody, serene, tense, warm palette, muted earth tones, high contrast.
- Quality and reference words - highly detailed, sharp focus, 35mm, cinematic. Use these sparingly; a few help, a pile of them cancels out.
Read those parts back to back and you get a natural sentence or a comma-separated stack. Both work. Commas give you cleaner control over what to add or remove between iterations.
Parameters, explained
Parameters go at the end of the prompt, each starting with a double hyphen. They are switches: they change how the model renders rather than what it renders. Here are the ones that earn their place.
| Parameter | What it does | Typical use |
|---|---|---|
--ar | Aspect ratio (width:height). | --ar 16:9 for widescreen, --ar 2:3 for a portrait, --ar 1:1 square. |
--v | Model version. | --v 6. Newer versions read prompts more literally and render finer detail. |
--style raw | Turns down Midjourney’s built-in aesthetic bias. | When you want the model to follow your words instead of prettifying them. Good for product and reference shots. |
--stylize (--s) | How strongly the house style is applied, 0 to 1000. | Low (--s 50) stays literal; high (--s 750) is more artistic and looser. |
--chaos (--c) | Variety across the four initial images, 0 to 100. | --c 30 for a spread of ideas; leave it off when you already know what you want. |
--q | How much rendering time and detail to spend. | --q 1 is standard; lower drafts faster, higher is not always better. |
--no | Excludes things from the image. | --no text, watermark, people. Name what you do not want; do not phrase it as "without". |
--seed | Fixes the random starting point. | Reuse the same --seed 12345 to keep composition stable while you tweak wording. |
A practical note on --no: writing "a table without clutter" often makes the model draw clutter, because it reads the noun. --no clutter is the reliable way to remove something.
Photorealism vs illustration vs 3D
The medium words you choose steer the whole render. Match the vocabulary to the look.
Photorealism
Borrow the language of photography. Name a camera or lens feel, a film stock, an aperture, and real lighting. "85mm portrait, f/1.8, shot on Portra 400, soft window light, shallow depth of field." Add --style raw and keep --stylize low so the model does not push it toward illustration.
Illustration
Name the technique and, if you like, an era or tradition: "flat vector illustration", "ink and watercolour", "1950s children’s book gouache", "bold line art with cel shading". Describe colour as a palette rather than as physical light. A higher --stylize value tends to help here.
3D and render
Use render vocabulary: "3D render", "octane render", "clay model", "isometric", "soft studio lighting", "subsurface scattering", "matte plastic". Mention the material and how light behaves on it. Isometric framing plus a plain background reads as a clean product-style 3D scene.
Word order and specificity
Put the thing that matters most first. If the subject is the point, lead with it; if the mood is the point, you can open with it. Everything after the subject refines it. Two habits pay off immediately:
- Trade adjectives for nouns. "a rusted 1970s diner" gives the model more to hold onto than "an old sad place".
- One idea per comma. Stacking "cinematic, dramatic, epic, moody, atmospheric" dilutes each word. Pick the two that carry the image.
Vague requests get generic results. Specific ones get the picture in your head. This is the same principle covered in how to write AI prompts, applied to images.
Six example prompts
Each of these is a working starting point. Change the nouns, keep the structure.
Cinematic portrait of an elderly fisherman, weathered face, wool sweater, standing on a harbour wall at dawn, soft golden light, 85mm, shallow depth of field, muted cool palette --ar 4:5 --v 6 --style raw --s 100Flat vector illustration of a cosy reading nook, potted plants, warm lamp glow, mustard and teal palette, clean shapes, subtle grain --ar 3:2 --v 6 --s 400Isometric 3D render of a tiny island cafe, pastel colours, soft studio lighting, matte materials, plain background, highly detailed --ar 1:1 --v 6 --no textMacro photograph of a dew-covered spiderweb at sunrise, backlit droplets, dark green bokeh background, sharp focus, natural light --ar 3:2 --v 6 --style raw --s 50Moody concept art of an abandoned space station corridor, flickering emergency lights, wet floor reflections, atmospheric haze, wide shot --ar 21:9 --v 6 --c 15Ink and watercolour illustration of a fox curled asleep under autumn leaves, loose brushwork, warm earthy tones, soft paper texture --ar 2:3 --v 6 --s 350 --no borderIterating with variations
The first four images are a draft, not the answer. Work in small, deliberate passes rather than rewriting the whole prompt each time.
- Generate, then pick the composition closest to your intent and request variations on it.
- Change one thing per round - swap the lighting, or the palette, or the lens - so you can tell what caused the change.
- Lock a
--seedonce you like the framing, then adjust wording to refine detail without the layout jumping around. - If results feel too polished or too samey, lower
--stylizeor raise--chaos; if they ignore your words, add--style rawand lower--stylize.
Keep the versions that almost worked. A prompt that missed on lighting but nailed the pose is one edit away from the shot.
Does this transfer to other models?
The structure does; the parameters do not. DALL·E, Flux, and Stable Diffusion all reward the same clear ordering of subject, detail, medium, lighting, and composition. What differs is the syntax around it. DALL·E prefers full descriptive sentences and has no --ar style flags. Stable Diffusion uses a separate negative prompt field instead of --no, and exposes its own controls (steps, CFG scale, sampler). Flux sits closer to natural language. So treat the six-part anatomy as portable and the double-hyphen switches as Midjourney-specific.
Want more finished prompts to study across tools? See our AI prompt examples, or build one for your exact scene with the free prompt generator.