How to Write Midjourney Prompts That Actually Work (V7 Guide, 2026)

TL;DR

In May 2026, Midjourney V7 prompts work meaningfully differently from V6:

1. --cref is gone in V7 — use Omni Reference instead for character consistency
2. --sref (style reference) still works and is the most powerful style tool in the platform
3. Natural language beats keyword stacking — “a woman reading in a sun-lit café corner, late afternoon light angling through tall windows” produces better results than “woman, reading, café, sunlight, warm, cozy”
4. Parameter discipline matters more than parameter quantity — --ar, --s, --chaos, --weird, --quality are the ones that earn their keep

The single biggest mistake in 2026 Midjourney prompts: dumping comma-separated keywords. V7 understands sentences. Write sentences.

What changed with V7

If you learned Midjourney on V5 or V6, two changes matter:

--cref is incompatible with V7

This is the most-broken pattern in 2026 prompts. --cref (character reference) was the V6 feature for consistent characters across generations — provide a reference image, get the same character back.

In V7, --cref is incompatible. If you use it, the system either errors or ignores the parameter entirely.

The replacement is Omni Reference — Midjourney’s unified system that blends character traits and style automatically. Drop a reference image into your prompt; V7 reads both subject identity and visual style from it as one signal.

--sref still works and is critical

--sref (style reference) takes either a URL of an image or a style code from Midjourney’s internal library and applies that style to your prompt:

[your prompt] --sref https://example.com/your-image.jpg

Or with Midjourney’s library codes:

[your prompt] --sref 1234567890

You can adjust the influence of the style with --sw (style weight, 0–1000):

[your prompt] --sref 1234567890 --sw 500

For brand consistency, recurring projects, or matching an existing visual identity, --sref is the most powerful tool in Midjourney V7.

The prompt structure that works

Every effective Midjourney V7 prompt follows the same skeleton:

[Subject] in [Environment], [Lighting/mood], [Style/medium] --[parameters]

Example:

A woman in a charcoal coat reading a paperback at a café table, late afternoon light through tall windows, 35mm film aesthetic --ar 3:2 --s 250

Each component has a job:

• Subject — what you’re depicting (be specific)
• Environment — where they are, what’s around them
• Lighting/mood — the deciding factor for atmosphere
• Style/medium — film, oil painting, digital illustration, watercolor, etc.
• Parameters — control over composition, style strength, variation

You can omit any component, but the prompts that consistently produce strong results include all four plus the right parameters.

Parameters that matter in V7

--ar (aspect ratio) — the most underrated parameter

Aspect ratio changes composition as much as it changes crop. A 16:9 landscape and a 4:5 portrait of the “same” prompt are dramatically different images:

• --ar 1:1 — square, default, balanced compositions
• --ar 4:5 — portrait social posts (Instagram)
• --ar 9:16 — vertical phone video / Reels / Shorts
• --ar 3:2 — classic photography aspect (35mm)
• --ar 16:9 — cinema, landscape banners
• --ar 21:9 — ultrawide cinematic

Pick the aspect ratio first. It influences everything else.

--s (stylize) — how literal vs. artistic

Range: 0–1000. Default is around 100.

• --s 0 — extremely literal interpretation of your prompt
• --s 100 — default, balanced
• --s 500 — more artistic, painterly, stylized
• --s 1000 — heavily stylized, prompt fidelity drops

For photorealism, lean low (--s 50). For concept art, lean high (--s 500). For most work, default (--s 100) is right.

--chaos — variation between results

Range: 0–100.

• --chaos 0 — the four returned images are similar variations of one composition
• --chaos 50 — meaningful variety in composition and approach
• --chaos 100 — wildly different interpretations of the prompt

For exploration (“show me different takes on this concept”), high chaos. For refinement (“I like this composition, give me variations”), low chaos.

--weird — surreal deviation

Range: 0–3000. New in recent V7 updates.

• --weird 0 — normal, expected output
• --weird 500 — slightly surreal, unexpected
• --weird 2000 — clearly surreal, dreamlike
• --weird 3000 — highly abstract, often unusable

For creative concepts, 250–750 is the sweet spot. Above 1500, results become unreliable.

--quality — render quality vs speed

Range: 0.25–2.

• --quality 0.5 — faster generation, lower detail
• --quality 1 — default, balanced
• --quality 2 — slowest, highest detail

Most work is fine at default. Use --q 2 for final output you’ll print or use commercially.

Prompts that work — examples

Cinematic photography

Prompt:

A man in a worn denim jacket walking past a neon-lit ramen shop in Tokyo at night, rain-slicked street, shallow depth of field, 35mm film aesthetic, Kodak Portra 400, melancholy mood --ar 16:9 --s 150

What’s working:

• Specific clothing detail (“worn denim jacket”)
• Specific setting (“neon-lit ramen shop in Tokyo”)
• Lighting + mood (“rain-slicked, shallow depth of field, melancholy”)
• Specific medium (“35mm film, Kodak Portra 400”)
• Aspect ratio matches cinematic intent

Concept art / illustration

Prompt:

A young woman with copper hair standing on a cliff overlooking a sea of bioluminescent jellyfish, twilight, swirling clouds catching the last light, painterly digital illustration, brushwork visible --ar 3:2 --s 500 --chaos 30

What’s working:

• Color signal (“copper hair”) gives the AI a hook
• Specific scene with stakes (“standing on a cliff overlooking…”)
• Light direction (“twilight, last light”)
• Medium specified (“painterly digital illustration, brushwork visible”)
• Higher --s (500) leans into painterly aesthetic
• Modest chaos (30) for variations on the same concept

Product photography

Prompt:

A minimalist white ceramic vase on a polished concrete shelf, single white tulip emerging, soft window light from the left, deep shadows on the right, magazine still-life photography --ar 4:5 --s 50 --q 2

What’s working:

• Concrete subject + concrete setting
• Light direction (“from the left”) + shadow placement
• Medium (“magazine still-life photography”)
• Low --s (50) for literal interpretation
• High --q (2) for printable detail
• Portrait aspect for Instagram-friendly product shot

Brand-consistent series

Once you’ve found a style you like:

[your new subject prompt] --sref [URL of your reference image] --sw 500

This produces a new image in the visual language of your reference. Repeat across many subjects to build a consistent series.

Common mistakes that ruin output

❌ Comma-separated keyword stacking

woman, reading, cafe, sunlight, warm, cozy, beautiful, professional, high quality, 8k, masterpiece

V7 doesn’t reward keyword stacking. Half those tokens are wasted. Prefer:

A woman reading in a sun-lit café corner, late afternoon light angling through tall windows

❌ “Quality booster” tokens

8k, masterpiece, award-winning, trending on ArtStation, ultra detailed — these helped early models. They don’t help V7. Use --q 2 for high quality if needed; skip the magic-word tokens.

❌ Negative prompts as “no X” in the prompt body

V7 has an explicit --no parameter:

[your prompt] --no text, signs, watermarks

Don’t write “no text in the image” inside the prompt — it can backfire. Use --no.

❌ Using --cref in V7 (still seen everywhere)

Old tutorials and 2024-era guides still recommend --cref. It doesn’t work in V7. Use Omni Reference (drop the image into the prompt) or --sref for style consistency.

❌ Vague subjects

“A beautiful image of a woman” gives you generic stock-photo output. Specificity is everything: hair color, clothing, environment, action, lighting. The more your prompt looks like a description in a screenplay, the better V7 handles it.

Practical workflow

For a single image

1. Write a structured prompt: subject + environment + lighting + style + parameters
2. Generate
3. Pick the closest result, use vary (subtle) or vary (region) to refine
4. Upscale the final winner

For a consistent series

1. Generate one image you love
2. Save its URL
3. Use it as --sref for every subsequent image in the series
4. Adjust --sw (style weight) higher for stronger consistency, lower for variation

For exploration

1. Write a moderately specific prompt
2. Set --chaos 50 or higher
3. Generate multiple times to see different interpretations
4. Pick a direction; lower chaos and refine

What’s coming

• V8 is rumored for late 2026, with stronger video features and possibly a return of dedicated --cref (Omni Reference is convenient but less granular)
• Conversational editing (à la ChatGPT Images 2.0) is the trend pulling Midjourney forward — expect dialogue-driven editing features in the next major version

For more, see Midjourney vs DALL-E (now ChatGPT Images 2.0), Best AI for product mockups (forthcoming), and How AI image generators actually work.

One closing tip

The biggest leap in your Midjourney work won’t come from learning more parameters. It’ll come from describing your subject more specifically. Most prompts fail because the writer hasn’t decided what they actually want to see. The 30 seconds you spend choosing between “a woman” and “a woman in her early 40s with copper hair, gold-rimmed glasses, slightly amused expression” buys more output quality than any parameter tweak.

Decide first. Then describe.