✓ Updated July 5, 2026
Sources: Midjourney official docs · Parameter List v6 · Community testing
How to Prompt Midjourney: Complete Guide from Beginner to Expert (2026)
Midjourney is not a chatbot. It is a text-to-image model with its own language — and the gap between a beginner prompt and an expert prompt is enormous. A beginner gets a generic image. An expert gets exactly what they visualized. The difference is understanding how Midjourney weights words, uses parameters, and responds to composition, lighting, and style descriptors. This guide takes you through all of it.
The Core Insight
Midjourney is not reading your prompt like a sentence — it is weighting individual words and phrases against a massive image dataset. The order of your words matters (earlier words have more weight), specific descriptors outperform vague ones, and parameters like --ar, --chaos, --style, and --sref give you precise control that words alone cannot. Master these and you control the output.
AI Prompting Guides — Full Series
🟣
Claude
XML tags, system prompts, 30% placement rule
🟢
ChatGPT
Output contracts, reasoning effort, few-shot
🔵
Gemini
Search grounding, 1M context, multimodal
🎨
Midjourney
Parameters, weighting, style references
⌨️
Cursor
.cursorrules, @mentions, agent mode scope
⚡
Grok
Live X data, DeepSearch, Imagine prompts
Prompt Anatomy — The Structure That Works
Basic level
How Midjourney reads your words
Midjourney does not understand sentences the way Claude or ChatGPT do. It tokenizes your prompt and assigns attention weights to each word and phrase. Words at the beginning of your prompt receive more weight than words at the end. This means the most important element of your image — the subject — should come first, every time.
Long, grammatically correct sentences often produce worse results than structured keyword phrases. Midjourney was trained on image-caption pairs, not prose. It responds better to descriptive noun phrases than to instructions like "please create" or "I want."
The 5-part prompt structure
| Position | What goes here | Why it matters |
|---|
| 1. Subject | The main thing in the image | Highest attention weight — always first |
| 2. Style / medium | Photography, oil painting, 3D render, illustration | Sets the visual language of the entire image |
| 3. Composition | Close-up, aerial view, wide shot, centered | Controls framing and spatial relationship |
| 4. Lighting | Golden hour, studio lighting, neon, overcast | The single biggest impact on mood and quality |
| 5. Parameters | --ar 16:9 --chaos 20 --style raw | Precise technical control, always at the end |
✗ Weak — no structure, generic
a woman in a forest
✓ Strong — structured, specific, controlled
portrait of a woman standing in an ancient redwood forest, analog film photography, Kodak Portra 400, shallow depth of field, dappled golden hour light filtering through canopy, mist at ground level, looking away from camera, editorial style --ar 2:3 --style raw --chaos 5
Word order and weight
The first 3–5 words of your prompt carry the most influence. Put your subject and its most important characteristic first. If you want "a dramatic storm over a medieval castle," write exactly that — not "a medieval castle with a dramatic storm above it." The storm being mentioned first changes what Midjourney emphasizes.
Test this yourself: Run "cat portrait, oil painting" and "oil painting, cat portrait." You will get different results. The first produces a portrait-first image in an oil style. The second produces an oil painting aesthetic that happens to contain a cat portrait — subtly different emphasis.
Every Key Parameter Explained
Intermediate level
--ar [width]:[height]
Aspect ratio. Controls image dimensions. Default is 1:1 (square). Use 16:9 for widescreen, 9:16 for portrait/mobile, 2:3 for print portrait, 3:2 for landscape photography. This changes not just cropping but composition — Midjourney adapts the scene to the ratio.
Examples: --ar 16:9 (cinematic) | --ar 9:16 (Instagram story) | --ar 4:5 (Instagram post) | --ar 2:3 (book cover)
--chaos [0–100]
Controls variety across the 4-image grid. At 0 (default), all 4 images are similar and closely follow the prompt. At 100, the 4 images are wildly different, often departing significantly from the prompt. Use low chaos (0–10) for consistent, reliable generation. Use high chaos (40–80) for exploration and unexpected directions. Never use 100 for production work — you cannot predict what you get.
--chaos 0 (consistent) | --chaos 20 (slight variation) | --chaos 60 (exploratory)
--style raw
Disables Midjourney's automatic aesthetic beautification. By default, MJ applies its own "taste" — increasing saturation, smoothing details, adding a signature look. --style raw gives you more literal prompt interpretation with less AI flavor. Use it when: you want photorealistic outputs, you are working with a reference image, or you want precise control over the final look without MJ imposing its aesthetic.
Use with: photography prompts, product shots, architectural renders, portrait work
--stylize [0–1000]
Controls how strongly Midjourney applies its trained aesthetic style. Default is 100. Higher values (250–750) produce more "beautiful" but less literally accurate images — MJ takes creative liberties. Lower values (0–50) stick closer to the prompt at the cost of some visual polish. Think of it as the dial between "exactly what I said" and "MJ's artistic interpretation of what I said."
--stylize 50 (literal) | --stylize 100 (default) | --stylize 500 (strongly stylized)
--no [elements]
Negative prompting. Tells Midjourney what to exclude from the image. More reliable than trying to avoid things through positive phrasing. Use when you keep getting unwanted elements. Can list multiple items: --no text, watermarks, blur, lens flare
--no text --no people --no modern buildings --no blur
--seed [number]
Fixes the random seed used for generation. Two prompts with the same seed and similar wording will produce very similar images. Use this to: iterate on a composition you like, reproduce a result you generated before, or run A/B tests where only one element of the prompt changes.
--seed 42 (any number from 0–4294967295)
--weird [0–3000]
Adds unconventional, experimental aesthetics. The model deliberately pursues unexpected interpretations. Low values (50–200) add interesting quirks. High values (1000+) produce genuinely strange, sometimes incoherent results. Use for creative exploration, concept art, and editorial imagery where surprising outputs are welcome.
--weird 100 (subtle) | --weird 500 (noticeably unusual) | --weird 2000 (surreal)
Style, Lighting, and Composition Vocabulary
Intermediate level
Photography style descriptors that work
| Descriptor | What it produces |
|---|
analog film photography | Film grain, slight color shift, vintage feel |
Kodak Portra 400 | Warm skin tones, soft shadows, specific film stock look |
Fuji Velvia 50 | High saturation, deep greens, landscape photography feel |
35mm photography | Natural perspective, documentary feel |
medium format photography | Exceptional detail, slight compression, editorial look |
macro photography | Extreme close-up, shallow depth of field |
drone photography, aerial view | Bird's eye perspective, wide establishing shot |
Lighting — the most impactful single word
| Lighting descriptor | Effect | Best for |
|---|
golden hour lighting | Warm, directional, long shadows | Portraits, landscapes, emotional scenes |
blue hour lighting | Cool, soft, twilight atmosphere | Urban, melancholic, cinematic |
studio lighting, three-point lighting | Clean, professional, no harsh shadows | Product shots, headshots |
neon lighting | Vivid color, urban glow, cyberpunk feel | Night scenes, editorial |
overcast diffused lighting | Flat, even, no shadows | Fashion, product photography |
chiaroscuro lighting | Extreme contrast, dramatic shadows | Portrait, fine art, drama |
bioluminescent lighting | Glowing, otherworldly blue-green | Fantasy, sci-fi, nature |
Art style descriptors
| Descriptor | What it produces |
|---|
oil painting, impasto technique | Thick visible brushstrokes, classic fine art |
watercolor illustration | Soft edges, translucent color layers |
isometric 3D render | Game-like perspective, clean geometry |
concept art, matte painting | Professional illustration, fantasy/sci-fi environments |
editorial illustration, flat design | Clean vectors, modern publication look |
ukiyo-e woodblock print | Japanese traditional art style, bold outlines |
hyperrealistic, photorealistic, 8K | Maximum detail and realism |
Advanced: Style References, Character References, Weighting
Advanced level
Style references (--sref) — the consistency game-changer
Style references let you provide an image URL and have Midjourney extract and apply its visual style to your prompt. This is how professional designers maintain visual consistency across a project — all images share the same color palette, rendering style, and aesthetic mood.
a bustling Tokyo street market at night --sref [URL of reference image] --sref [second URL] --ar 16:9 --style raw
You can use multiple --sref URLs to blend styles. The model averages the visual characteristics across all references. Weighting works here too: --sref URL1::2 URL2::1 makes URL1 twice as influential as URL2.
Finding good sref sources: Use your own previous Midjourney outputs (right-click → copy image address), art reference sites, or stock photo URLs. Brand photography, film stills, and fine art works well. Avoid images with strong subjects — you want style, not subject matter, to transfer.
Character references (--cref) — consistent characters across images
Character references let you maintain a consistent person or character across multiple images. Provide an image URL of your character, and Midjourney preserves their appearance (face, hair, general look) while changing pose, setting, and action.
the same woman from the reference, now running through a rain-soaked alley, cinematic, neon reflections --cref [character image URL] --cw 100 --ar 16:9
--cw controls character weight (0–100). Higher values maintain more fidelity to the reference character at the cost of prompt flexibility. Lower values allow more creative interpretation while keeping general resemblance.
Prompt weighting — controlling emphasis
Use :: to assign relative weight to parts of your prompt. This is how you tell Midjourney which element matters most when there is potential conflict.
ancient castle ::3 storm clouds ::2 lightning ::1 --ar 16:9
This means: the castle gets 3 units of emphasis, storm clouds get 2, lightning gets 1. The castle will dominate the composition. Without weights, Midjourney distributes emphasis based on word order and its own interpretation.
Negative weighting (::-1) suppresses elements, similar to --no but more granular:
forest scene, mysterious atmosphere ::2 people ::-0.5 --ar 3:2
Iteration strategy — how professionals work
Experts do not write one prompt and hope. They iterate systematically:
| Stage | Goal | Technique |
|---|
| 1. Explore | Find the right direction | High chaos (40–60), loose prompt, 4 grids |
| 2. Lock composition | Fix the layout you want | Use Vary (Subtle) on your best result |
| 3. Refine style | Get the exact look | Add style/lighting descriptors, lower chaos |
| 4. Lock seed | Reproduce consistently | Copy seed from job ID, add --seed |
| 5. Upscale | Final output | U1–U4 buttons, then Upscale (2x/4x) if needed |
Before/After Rewrites — Real Examples
Example 1 — Product photography
✗ Weak
a perfume bottle on a table
✓ Strong
luxury glass perfume bottle with gold cap, product photography, studio lighting, white marble surface, soft shadow, shallow depth of field, high-end fashion editorial style, clean minimal background, extreme detail --ar 4:5 --style raw --stylize 200
Example 2 — Portrait
✗ Weak
portrait of an old man
✓ Strong
close-up portrait of an elderly fisherman, weathered face, deep wrinkles, kind eyes, analog film photography, Kodak Tri-X 400 black and white, late afternoon directional light, catching in crow's feet, character study, documentary photography --ar 2:3 --style raw --chaos 5
Example 3 — Environment / scene
✗ Weak
futuristic city
✓ Strong
vast megacity at blue hour, elevated highway systems weaving between glass and carbon fiber towers, bioluminescent plants integrated into architecture, hovering transit pods leaving light trails, street level market stalls with warm lanterns contrasting cold tower lights, concept art, matte painting, cinematic wide shot, extreme detail --ar 21:9 --stylize 750 --chaos 15
5 Failure Modes and Fixes
Failure mode 01
Image looks generic despite a detailed prompt
Root cause: The prompt contains specific subjects but generic style and lighting descriptors. "A woman in a park" with detailed clothing but no lighting or medium produces a competent but unremarkable result — MJ defaults to its average aesthetic.
Fix: Add medium + lighting as early as possible. "Analog film photography, golden hour, [subject]" immediately differentiates the output. The lighting descriptor alone changes the entire mood of the image.
Failure mode 02
Text in the image is garbled or unreadable
Root cause: Midjourney is an image model, not a text renderer. It generates shapes that look like text but are not real words. Earlier versions were especially bad at this — V6 improved significantly but still struggles with more than 3–4 words.
Fix: For short text (signs, logos, labels), use quotation marks: a neon sign reading "OPEN 24H" in a rainy alley. For longer text or precise typography, generate the image without text in Midjourney, then add text in Photoshop or Canva. Add --no text, writing to suppress unwanted text that appears uninstructed.
Failure mode 03
Hands and fingers look wrong
Root cause: Human hand anatomy is notoriously difficult for image models. Finger count, joint positions, and proportions frequently fail — especially at small image sizes or when hands are not the primary focus.
Fix: Compose shots where hands are not prominent. Use framing that cuts off at the wrist. If hands must be shown, upscale to maximum resolution, then fix in Photoshop using the generative fill tool. Adding "perfect hands, anatomically correct hands" to the prompt helps marginally but is not a reliable fix.
Failure mode 04
The 4-image grid is too similar — no useful variation
Root cause: Low chaos value (0 is default) plus a very specific prompt produces near-identical results across all 4 grid images. You are not getting options, you are getting 4 copies.
Fix: Raise --chaos to 20–40 for meaningful variation while keeping outputs relevant. Alternatively, run the same prompt 3–4 times without a fixed seed — natural sampling variation produces different results each run.
Failure mode 05
Style reference transfers the subject, not just the style
Root cause: When using --sref with an image that has a strong, distinctive subject (a specific person, landmark, or object), Midjourney extracts both the style AND subject characteristics. Your intended subject gets mixed with the reference subject.
Fix: Use abstract, texture-rich reference images for sref rather than images with prominent subjects. Landscapes, textures, color palette images, and abstract art transfer style cleanly without bleeding subject matter. Reduce --sw (style weight) to dial down the reference influence: --sref URL --sw 50
Midjourney vs DALL-E: Which Wins?
Full side-by-side comparison on image quality, consistency, and use cases.
Midjourney vs DALL-E →