Midjourney v6 Prompt Craft: A Field Guide

TL;DR — v6 broke v5’s habits: stop stacking junk keywords (“8k, octane render, trending on artstation”) and describe one coherent scene in plain language [1][2]. Front-load the subject, keep prompts ~50–150 words, and reach for parameters (--stylize, --chaos, --style raw) and references (:: weights, --sref, --cref) only when the words alone can’t get there [8][15]. ⚠ v6 is a legacy model in 2026 (default has moved through v7 to v8 [40]), but its core craft — describe what you do want, concisely — still transfers [34][36].

What actually changed in v6

Midjourney’s own v6 announcement was blunt: prompting is “significantly different than V5,” users must “relearn” it, and the model is “MUCH more sensitive” to wording while being far better at understanding explicit descriptions [1]. The practical effect is a paradigm shift from comma-separated keyword salad toward natural language — write a full descriptive sentence, the way you’d brief a photographer on a scene [5].

Three concrete consequences:

• Junk tokens stopped helping. Camera and quality boilerplate (“8k, best quality, Canon EOS R5”) can be stripped with nothing lost, because v6 understands prompts well enough not to need them [2][9].
• Longer, coherent prompts hold up. v6 takes long, conversational prompts and fulfills nearly every element — a marked upgrade over v5.2’s comprehension [3]. In token-retention testing it held all prompt elements through seven diffusion steps before dropping any, one step further than v5.2 [4].
• Literalness levers appeared. Text inside "double quotes" is rendered as a literal instruction, and --style raw produces a more literal, less auto-beautified result — recommended when you need legible lettering [6].

Anatomy of a v6 prompt

Midjourney’s docs frame a prompt around seven considerations — subject, medium, environment, lighting, color, mood, composition — and stress that short, simple prompts usually generate the best images (brevity buys variety, length buys control) [7]. A widely-shared working order:

[subject] · [style/medium] · [lighting] · [color palette] · [composition] · [mood] --ar [ratio] --style raw

[8] A parallel six-part recipe runs style → subject → setting → composition → lighting → details [10].

Three rules that matter more than the exact template:

One nuance: word order’s effect is muted when a strong artist reference anchors the aesthetic, but becomes decisive for weak or undefined styles [8].

Parameters that move the needle

Parameters do two jobs: steer the model’s aesthetic latitude, and shape output geometry. Ranges and defaults below are from the official docs.

⚠ High --stylize washes out fidelity. At --s 750 Midjourney takes creative liberty and may ignore some specified details to polish the image — so if v6 is “disobeying” your prompt, lower the stylize before adding more words [31].

References and weights

The double-colon :: syntax is the foundation of fine control. concept:: number weights a concept relative to others — space::2 ship makes “space” twice as influential as “ship”. No space before ::, one space after [15]. v6 accepts decimal weights for fine tuning [21].

Negation is just a negative weight. --no X is exactly shorthand for X::-0.5 [15]. Weights are relative (wood::1 teapot::-1 ≡ wood::10 teapot::-10), but the sum of all weights must stay positive or Midjourney errors [21].

--sref random resolves to a reusable numeric code shown in the prompt after generation — a free way to discover and lock aesthetics [24]. --cref is precision-limited: it won’t reproduce exact dimples, freckles or t-shirt logos, and is not designed for real people or photos [25].

What backfires (the skeptic’s section)

The single biggest failure mode is carrying v5 habits into v6 [26].

• Keyword stacking averages to mush. Pile on five-plus style modifiers and they blend into a muddy average rather than compounding — exactly why “highly detailed, 8k, octane render, trending on artstation” lost its punch [29]. Vague quality boilerplate (“award-winning,” “photorealistic,” “4k/8k”) dilutes the prompt [9].
• Long prompts silently drop tokens. The longer the prompt, the more likely Midjourney assigns weight 0.00 to important keywords — especially late ones — and ignores them [30]. v6 begins omitting detail past a token threshold; fusing words with underscores into a single token (stained_glass) raises per-token weight [33].
• Diagnose with /shorten. It scores every token 0–1 and strikes through near-zero words, making your dead filler visible [32].
• Negation in the positive prompt backfires. Writing “without hats” forces the model to parse a negation and can emphasize the very thing you wanted gone; the dedicated --no hat removes it reliably [27][28].

The consistent, official-aligned rule: describe what you DO want, concisely — not what to avoid.

Where v6 sits in 2026

v6’s lineage moved fast: v6.1 became default on July 30 2024, and v7 (released April 3 2025) took over as default on June 17 2025 [34]; by early 2026 Midjourney had shipped v8 on a rewritten engine with native 2K output [40]. v6 is still selectable via --v 6, but it’s two-plus generations behind.

For craft, the trajectory matters more than the version number. Midjourney’s own v7 note warns it’s “an entirely new model” that “may require different styles of prompting” and is “much smarter with text prompts” [35]. v7 improved prompt understanding ~35%, so simpler prompts now beat the parameter-heavy scaffolding v6 rewarded [36][37]. The direction of travel — less prompt-engineering, more plain description — is the same lesson v6 first taught.

How the prompting philosophies compare

Flux’s strict instruction-following is the cleanest contrast to Midjourney’s looser, taste-driven interpretation — ask Midjourney for three people and you may get five [39]. If you need exact counts, positions or text, that gap is the reason to reach for a rival; if you want a striking image fast, it’s Midjourney’s whole appeal.