Midjourney v6 · Prompt Playbook

Write scenes,
not keywords.

v6 broke v5's habits. Natural language replaced keyword soup — this is the field manual for what that actually means in practice. ^[1]

40 sources 8 min read expedition depth ⚠ v6 is legacy in 2026 — v8 is current

The v5 → v6 Paradigm Shift

Unlearn before you relearn

❌ v5 habit — kill this

portrait of woman, 8k, photorealistic, octane render, best quality, trending on artstation, highly detailed, Canon EOS R5, sharp focus, bokeh

Junk tokens blend into mush. v6 assigns weight 0.00 to filler — "8k", "best quality", camera models — and silently ignores them. Stacking 5+ style modifiers averages them into blandness. ^[2]^[29]

✓ v6 approach — do this

a 28-year-old woman standing in golden-hour sunlight, candid street portrait, warm amber light catching her hair, shallow depth of field --ar 4:5 --style raw

Describe what you see in the frame. v6 is much more sensitive to specific wording and rewards explicit physical, contextual, and emotional detail over quality boilerplate. ^[9]^[3]

Anatomy of a Prompt

Front-load what matters most ^[7]

Subject

who/what
leads here

Medium

oil paint
photograph
3D render

Environment

where
background
setting

Lighting

golden hour
studio rim
neon glow

Color

warm amber
desaturated
high contrast

Mood

melancholic
tense
serene

Composition

wide angle
close-up
bird's eye

[subject with qualifiers] · [medium/style] · [lighting] · [color palette] · [composition] · [mood] --ar 16:9 --style raw

Front-loading CRITICAL

Attention is distributed across the prompt. Earlier terms carry more visual weight — lead with the most important element, not "beautiful" or style preamble. Word order matters less when a strong artist reference anchors the aesthetic. ^[8]

Length sweet spot 50–150W

Short prompts buy variety; long prompts buy control. Past ~150 words, Midjourney starts assigning weight 0.00 to late keywords. v6 held all elements through 7 steps before dropping any — one better than v5.2. ^[30]^[4]

Specificity wins

"A dragon" → generic. Qualified subject with scales, pose, environment, expression → distinctive. v6 rewards explicit physical + contextual + emotional detail over generic quality signals. ^[9]

Parameters That Move the Needle

Reach for these when words alone can't get there

--stylize-s

Master "artistic freedom" slider. Low = tracks your prompt literally. High = Midjourney takes creative liberty and may ignore specified details to polish the image.

0 literal1000 flair

Default: 100

⚠ At --s 750+ MJ may ignore your specific details. If v6 "disobeys," lower stylize before adding more words.

^[11] ^[31]

--chaos-c

Controls variety across the four-image grid. High chaos produces images that diverge substantially from each other — and from the prompt. Good for exploring; bad for precision.

0 uniform100 wild

Default: 0

^[12]

--style raw

Disables MJ's auto-beautification. Produces more realistic, photo-like results. Required for legible in-image text generation. Compatible from v5.1+.

off (default)on = literal

Default: off

^[14]

--weird-w

Adds experimental, unconventional aesthetic qualities. Values above ~500 produce genuinely strange imagery. Experimental feature with known seed-locking issues.

0 normal3000 alien

Default: 0

⚠ Experimental — not fully compatible with --seed.

^[13]

--quality-q

Controls GPU time → detail. Only affects the initial four-image grid, not variations, upscales, or inpainting. 0.25 and 0.5 are faster for exploration.

0.25 fast1 full

Default: 1 · values: 0.25 / 0.5 / 1

^[19]

--seed

Sets the starting noise grid for reproducibility. Useful for iterating on a composition across prompt edits. Has the least impact on the final image of any parameter.

0random default4,294,967,295

Default: random

⚠ Turbo mode breaks reliable seed locking.

^[17]

Reference Syntax

Weights, style locks, character consistency

:: weights

Control the relative influence of concepts within a single prompt. Foundation of all fine-grained control.

space::2 ship
→ space 2× as influential as ship

wood::1 teapot::-1
→ ≡ wood::10 teapot::-10 (relative)

--no hat ≡ hat::-0.5

No space before ::, one space after ^[15]
Sum of all weights must be positive ^[21]
Decimal weights fine-tune: ::0.3

--sref

Transfer an aesthetic from a reference image or code. Controls style without locking composition.

--sref https://img.url/style.jpg
--sref 1234567890  # numeric code
--sref random    # discover + lock

--sw 65   # diluted (sweet spot)
--sw 175  # dominant style
--sw 0    # cancels --sref

--sw default: 100 · range: 0–1000 ^[23]
Sweet spot: 65–175 ^[24]
--sref random → reusable numeric code

--cref

Maintain a consistent character appearance across multiple generations.

--cref https://img.url/char.jpg

--cw 100  # face + hair + clothes
--cw 50   # face + hair only
--cw 0    # face only (swap outfit)

--cw default: 100 ^[25]
⚠ Won't copy exact dimples, freckles, logos
⚠ Not designed for real people or photos
Image weight --iw: 0–3, default 1 ^[22]

The Four Traps

Most failures trace to these

❌ Trap 1 — Keyword stacking

highly detailed, 8k, octane render, photorealistic, award-winning, trending on artstation, best quality

Five+ style modifiers average into mush. v6 assigns near-zero weight to boilerplate — they actively dilute real intent. ^[29]

✓ Fix — describe one coherent scene

worn leather journal on a rain-spattered café table, warm tungsten light, cinematic close-up, 1970s film grain

Specific physical and contextual detail. v6 holds all elements through ~7 prompt steps. ^[4]

❌ Trap 2 — Negative in positive prompt

a portrait without hats, no glasses, not wearing a jacket

Forces the model to parse negations — and can emphasize the very things you wanted removed. ^[28]

✓ Fix — use --no

a portrait --no hat glasses jacket

--no X is shorthand for X::-0.5. Describe what you DO want; exclude with --no. ^[27]

❌ Trap 3 — High stylize ignoring the prompt

detailed scene with specific props and costume --s 750

At --s 750, MJ takes so much creative liberty it may ignore specific details to produce what it considers a "better" image. ^[31]

✓ Fix — lower stylize first

detailed scene with specific props and costume --s 100

If v6 "disobeys," lower --stylize before adding more descriptive words. ^[31]

❌ Trap 4 — Late-token burial

a beautiful serene landscape [200 more words...] with a red barn in foreground

Tokens late in a long prompt get weight 0.00 and are silently ignored. The red barn appears in 0% of outputs. ^[30]

✓ Fix — front-load or fuse

red_barn in foreground, serene landscape [remaining detail...]

Front-load critical elements. Or fuse as red_barn — one token, full weight. ^[33]

Signal Boosters & Diagnostics

When prompts underperform

⌖

/shorten

Scores every token 0–1 and strikes through near-zero-weight words. Run this on any underperforming prompt — you'll see which words cost you and give nothing back. ^[32]

/shorten highly detailed 8k portrait of a woman in sunlight
→ ~~highly detailed~~ ~~8k~~ portrait woman sunlight

_

Underscore fusion

MJ tokenizes on whitespace. Fusing words with underscores makes them a single token — raising that token's effective weight. Critical for compound concepts. ^[33]

stained glass  → 2 tokens (each ~0.5 weight)
stained_glass  → 1 token  (full weight)
golden_hour, art_nouveau, ink_wash

Quoted literals

Text in double quotes is treated as a literal instruction — specifically trained in v6. Combine with --style raw for the best in-image text. ^[6]

a neon sign reading "OPEN 24HRS" --style raw
→ legible text, not decorative swirls

↔

--ar quick ref

Set at prompt end, no space before dashes. Composition strategy should align with ratio. ^[20]

--ar 16:9   landscape / cinematic
--ar 4:5    portrait / feed-ready
--ar 1:1    square / cover art
--ar 2:3    book cover / poster

Where v6 Sits in 2026

Legacy model — craft transfers forward

Dec 2023

v6.0

Overhauled prompting. Natural language replaces keyword soup. "Relearn how to prompt."

Jul 30, 2024

v6.1

Became default. Better understanding of intent. This guide's primary target. ^[34]

Jun 17, 2025

New default. Prompt understanding +35%. Simpler prompts beat parameter-heavy approach. ^[36]

Mar 2026

Rewritten engine. Native 2K output. 5× faster. Current default. ^[40]

v6 is still accessible via --v 6 but is two generations behind. The direction of travel — less prompt-engineering overhead, more plain description — is the same lesson v6 first taught. v7's +35% prompt understanding made the parameter scaffolding v6 rewarded mostly unnecessary. ^[37]^[35]

Model	Prompting style	Adherence
Midjourney v6	Terse aesthetic phrases + parameters & refs	Aesthetic-first; may ignore literal specs ^[38]
DALL-E 3	Verbose natural-language sentences (ChatGPT rewrites)	High, conversational ^[38]
Flux	Photographic / camera terminology	Precise — follows instructions literally ^[39]
Stable Diffusion XL	Weighted (parentheses) keywords + negative prompt field	Keyword-control driven ^[38]

← Midjourney v6 Prompt Craft: A Field Guide · Diffusion Models expedition · Atlas

40 sources · expedition depth · 2026-06-09

Write scenes,not keywords.