Text-to-Image Prompt Weighting: A Practical Guide for Midjourney and Flux

Prompt weighting is not universal. Paste (beautiful:1.4) into Stable Diffusion and the model amplifies "beautiful." Paste the same string into FLUX and you get literal parentheses in the render — or nothing changes at all.

Midjourney v6 used cat::2 dog::1 to force composition. Midjourney v7 dropped inline :: multi-prompts in favor of natural language plus parameter weights (--ow, --sw, --stylize). FLUX never had :: — it weights by position in the sentence and guidance scale, not SD-style parentheses.

This guide maps what actually works on Midjourney and FLUX in 2026, when to roll back to v6 syntax, and why copying SD prompt packs into either model wastes tokens.

Three weighting systems (don't mix them)

| System | Models | Syntax |

| Token multipliers | SDXL, SD 1.5, Pony | (word:1.2), ((word)), [word] down-weight |

| Promptlet weights | Midjourney v1–v6.1 | concept A::2 concept B::1 |

| Positional + params | Midjourney v7, FLUX | Word order, --ow, --sw, guidance scale |

SDXL (keyword:weight) is covered in our dedicated SDXL guide. This article is Midjourney + FLUX only.

Midjourney v7: what replaced :: weights

Official Midjourney docs and API references (Evolink, Aliyun marketplace guides, mid-2026) state: v7 does not support inline multi-prompt `::` separators. The :: that still works lives inside parameters — e.g. --sref url1::2 url2::1 — not in the main text prompt.

v7 weighting levers:

1. Word order and sentence structure — subject and priority adjectives first
2. `--no` — exclude concepts (≈ -0.5 weight in v6 multi-prompt math)
3. Reference weights — --ow, --sw, --iw
4. `--stylize` — global aesthetic pressure vs literal prompt obedience

v7 text prompt: order = weight

Weak (everything equal):

forest, temple, mist, moonlight, ancient, cinematic

Strong (hierarchy without ::):

Ancient stone temple dominates the composition, rising from a misty forest. Soft moonlight filters through trees. Cinematic wide establishing shot.

First sentence = hero concept. Second = environment. Third = camera. Same four nouns, explicit hierarchy — no :: needed.

--no as negative weight

In v6, --no red equaled red::-0.5 in multi-prompt math. v7 keeps --no:

vibrant tulip field, spring morning, soft diffused light --no red flowers, shadows, watermark --v 7

Stack exclusions: --no text, logo, extra fingers, blurry background

Prefer positive framing when --no fights the scene: "white seamless background" beats --no clutter alone.

Parameter weights (v7 reference stack)

| Parameter | Range | Default | Controls |

| --ow | 1–1000 | 100 | Omni Reference (--oref) fidelity |

| --sw | 0–1000 | 100 | Style reference (--sref) strength |

| --iw | 0–3 | 1 | Image prompt vs text prompt balance |

| --stylize | 0–1000 | 100 | MJ house aesthetic vs raw obedience |

| --style raw | flag | off | Lowers stylize effect; prompt words dominate |

Competition rule (Aliyun MJ guide 2026): --ow, --sw, --iw, and --stylize fight for influence. High --stylize (250+) without raising --ow/--sw → character or style drift.

Multi-sref weighting (v7 — :: works here):

[prompt] --sref styleA.jpg::2 styleB.jpg::1 --sw 150 --v 7

styleA gets 2× the style influence of styleB within the sref parameter.

v7 worked example: balancing text vs reference

Goal: photoreal portrait, strong face lock, subtle style grade.

Woman in olive field jacket, three-quarter portrait, soft window light camera left, shallow depth of field --oref [URL] --ow 500 --sref [grade_url] --sw 60 --style raw --stylize 80 --v 7

• --ow 500 — face wins over scene novelty
• --sw 60 — color grade whispers, doesn't repaint
• --style raw --stylize 80 — text lighting description respected

If face drifts: raise --ow. If image looks over-processed: lower --sw and --stylize.

Midjourney v6: :: multi-prompt weights (still valid)

Need explicit concept blending with numeric control? Use v6:

ancient temple::3 misty forest::2 moonlight::1 --v 6.1

Each segment is a promptlet processed separately, then blended. Higher number = more canvas share.

v6 weight rules (official Midjourney docs)

• Default weight per promptlet: 1 if omitted
• v4–v6.1: decimals allowed — space::1.5 ship::0.8
• Negative weights shrink concepts: still life:: fruit::-0.5 (total must stay positive)
• --no word ≈ word::-0.5
• Final promptlet trailing :: is optional

When v6 :: beats v7 prose

• Deliberate concept collision — "space::2 ship::1" (boat in space, not spaceship)
• Style fusion with numeric ratio — art nouveau::2 cyberpunk::1
• Suppressing a word without --no — portrait:: background::-0.3

Discord still runs v6 multi-prompts; MJ Alpha website may not — check your interface (GeekyCuriosity / community notes, 2026).

v6 full stack example

space::2 ship::1 nebula background, cinematic --v 6.1 --ar 16:9 --stylize 150

Compare v7 equivalent (no ::):

Small wooden sailing ship in deep space, nebula background dominates, ship secondary in frame, cinematic --no sci-fi spaceship --v 7 --ar 16:9

FLUX: positional weighting (no ::, no (word:1.5))

Black Forest Labs trained FLUX on natural language captions — not booru tag soup. ImageToPrompt, fal.ai, and AI Wiki converge (2026):

• `(word:1.5)` is ignored — treated as literal text or no-op
• `++` / nested `((()))` emphasis — SD habits; no effect on FLUX.1 Dev/Pro
• Comma tag lists without grammar — weak vs full sentences
• Midjourney flags (--ar, --v 7) — render as literal text in the image

FLUX weighting = where you say it + guidance scale + reference conditioning.

FLUX priority stack (Subject → Action → Environment → Light → Style)

Template:

[SUBJECT — most detail here]. [ACTION/POSE]. [ENVIRONMENT]. [LIGHTING]. [STYLE/MEDIUM/CAMERA].

Example — emphasis on subject, not background:

A weathered fisherman with deep crow's feet and salt-and-pepper beard, wearing a yellow rain slicker — he is mending nets on a wooden dock. Harbor boats and grey sky in the background. Overcast soft daylight. Documentary photograph, 85mm shallow depth of field.

Background gets one clause. Face gets three descriptors upfront. No parentheses required.

Natural language emphasis (replaces weights)

When you need FLUX to care about one detail, say so in words:

• "with particular focus on the texture of the leather jacket"
• "emphasizing the catchlight in her left eye"
• "the red umbrella should be the brightest element in the frame"
• "background softly blurred, subject tack sharp"

This outperforms (red umbrella:1.4) every time on FLUX.

FLUX guidance scale (CFG) — the hidden weight knob

Not in the prompt string — set in UI/API:

• Low (2–3.5): creative, loose prompt adherence; good for exploration
• Mid (3.5–4.5): default sweet spot for FLUX.1 Dev
• High (5–7): strict literalism; prompt words dominate; risk of overcooked contrast

Long prompt + high guidance = every clause gets enforced → muddy averages. Shorten prompt or lower guidance.

FLUX 2 nuance (2026)

BFL official guide: FLUX 2 attention handles weighting internally. Community tests show simple parentheses for moderate emphasis may still help at the margin — but sentence structure beats syntax. Don't stack (((concept:1.5))); rewrite the sentence instead.

FLUX 2 multi-reference: weight via image order and API strength, not prompt :::

"Character from image 1 wearing the outfit from image 2, standing in the environment from image 3."

See our character consistency guide for ref weights.

Side-by-side: same intent, two dialects

Intent: red dress more important than ballroom.

Midjourney v6:

woman in ballroom:: red dress::3 --v 6.1

Midjourney v7:

Woman wearing a vivid red silk dress, full-length view, ornate ballroom interior softly blurred behind her --no muted wardrobe --v 7

FLUX:

Woman in a vivid red silk dress that fills the foreground, standing in an ornate ballroom with chandeliers receding into soft blur.

Stable Diffusion XL (not FLUX — shown for contrast):

woman in ballroom, (red dress:1.4), ornate interior -- negative: muted colors

Translation mistakes (fix these)

| Habit from SD | Midjourney v7 | FLUX |

| (masterpiece:1.2) | Ignored / literal noise | Ignored / literal | Use in SDXL only |

| 1girl, solo, best quality tags | Works sometimes; not optimal | Weak — write sentences |

| cat::2 dog::1 | Broken on v7 — use v6 or rewrite | N/A — use word order |

| --no | ✅ works | N/A — positive "without X" |

| Long negative prompt block | --no a few terms | Describe what you want instead |

| Same prompt all models | Rewrite per dialect | Rewrite per dialect |

Debugging weight problems

Symptom: wrong subject dominates.

• MJ v7: move subject to first sentence; add --no [competing concept]
• MJ v6: raise promptlet weight hero::3 background::1
• FLUX: shorten background clause; add "foreground dominated by [subject]"

Symptom: reference face drifts (MJ).

• Raise --ow 100 → 450; lower --stylize; add --style raw

Symptom: style ref overpowers scene (MJ).

• Lower --sw to 50–80; reduce sref count

Symptom: FLUX ignores adjective.

• Adjective may be too late in prompt — move earlier
• Lower guidance if image looks over-processed; raise if too loose

Symptom: (token:1.3) did nothing (FLUX).

• Expected. Rewrite with natural emphasis or split into two sentences.

PromptMake workflow

/image → pick Midjourney v7 or FLUX target → paste SD prompt → Improve or Recreate rewrites into dialect-correct structure (no spurious (weights) or --ar on FLUX).

A/B test: same scene description, two outputs — MJ with --style raw vs FLUX with guidance 4.0. Compare which obeyed hierarchy.

3 free /image generations per day — use for dialect translation before paid MJ/FLUX credits.

Quick reference card

Midjourney v7: order + --no + --ow/--sw/--iw/--stylize. No inline ::.

Midjourney v6: concept::weight promptlets + full param stack.

FLUX: subject-first sentences + natural emphasis + guidance scale + reference images.

SDXL: (token:weight) + negatives — different article.

Frequently asked questions

Does :: work in Midjourney v7?

Not in the main prompt. Only inside --sref url::2 style parameter weighting.

Can I use (word:1.5) in FLUX?

No meaningful effect on FLUX.1. Use word order and descriptive emphasis.

--no vs negative weight in v6?

Roughly equivalent to -0.5 on that term. --no red ≈ red::-0.5 when other weights sum positive.

Best --stylize for prompt obedience?

Product/photo: --style raw --stylize 50–100. Illustration: 150–250. Above 400: raise --ow if using --oref.

FLUX or MJ for precise concept ratios?

MJ v6 :: for numeric blend. v7 and FLUX: rewrite with explicit hierarchy sentences.

Does word order matter on MJ v7?

Yes — early sentences carry more weight even without ::. Lead with the hero.

Related articles

Midjourney prompts 2026 — full v7 parameter reference

SDXL weights & negatives — (token:1.2) when you need SD control

DALL·E vs Midjourney vs Flux — dialect comparison overview

Character consistency prompts — --ow and FLUX ref weighting

Art style cheat sheet — style-first ordering

7 elements of visual prompts — structure before weights

Bottom line

Prompt weighting is model-specific grammar. Midjourney v7 dropped :: from text — use sentence hierarchy, --no, and reference parameter weights. FLUX never had parentheses weights — use position, plain-language emphasis, and guidance scale.

Stop paste-porting SD prompt packs. Write once in ideas, translate three ways: v7 prose, v6 promptlets if you need numeric blend, FLUX subject-first sentences. The weight is in the structure — not the punctuation.