How to Write Midjourney Prompts That Actually Work in 2026

Most Midjourney frustration isn't the model — it's fighting defaults you didn't know were on.

v7 understands natural language better than the keyword-stuffing era. It also still applies a "house aesthetic" unless you tell it not to. Prompts that worked in v5 (tag soup, "8k masterpiece ultra detailed") often produce worse results now. Prompts that work in 2026 look like brief art direction notes plus parameters at the end.

This is the practical v7 playbook: structure, parameters that matter, reference modes, iteration workflow, and the habits to drop.

The v7 mindset shift

Three things changed from older guides:

Sentences beat tags. v7 parses scene grammar: subject, environment, light, camera. Comma-separated keyword piles compete with each other.

Defaults fight photorealism. Without --style raw, "professional photograph" still gets subtle Midjourney polish — skin too perfect, lighting too cinematic-painterly.

References beat adjectives for consistency. --sref, --cref, and --oref lock look and identity better than repeating "same character, same style" in prose.

Legacy listicles on this site (midjourney-prompt-tips, midjourney-v7-guide) cover individual parameters. This article is the integrated "what to run daily" workflow.

Prompt structure that v7 respects

Default formula:

[Subject + action] + [environment] + [lighting] + [camera/lens/medium] + [optional mood] + [parameters]

Subject first — Midjourney still weights early tokens. Put the hero noun up front.

One register per prompt — don't mix "whimsical storybook" with `--style raw` product photography. Pick illustrative OR photographic.

Parameters last — always trailing, double-dash syntax, space-separated.

Example (editorial portrait):

Woman in her 40s seated by a window, rain on glass, soft overcast daylight, muted wardrobe, shot on 85mm lens shallow depth of field, editorial fashion photography --ar 4:5 --style raw --stylize 100 --v 7

Example (illustrated — no raw):

Fox librarian sorting books in a cozy treehouse library, warm lantern glow, whimsical children's book illustration, gouache texture, rich autumn palette --ar 3:2 --stylize 350 --v 7

--style raw: non-optional for photos

--style raw disables Midjourney's automatic beautification. The model follows your camera and lighting words literally instead of painting over them.

Use raw for:

• Product shots, portraits, architecture, food, fashion editorial
• Any time outputs look "too Midjourney" despite realistic words

Skip raw (or expect stylization) for:

• Fantasy illustration, painterly concept art, maximalist aesthetics

Pair raw with lower stylize — typically --stylize 50–150 for photos. Default 100 is fine; 600+ fights raw mode.

Video/manual animation in v7: community reports raw improves motion fidelity too — same accuracy principle.

Stylize, chaos, and when to touch them

--stylize (--s) — how much Midjourney adds its own taste.

• 0–50: literal, diagram-like obedience
• 100–150: sweet spot with --style raw for photos
• 250–400: illustrative, editorial art
• 600–1000: maximum "wow" Midjourney look — prompt becomes a suggestion

--chaos — variation across the 4-up grid.

• 0–15: client work, consistent portraits, product variants
• 30–50: exploration while staying on-brief
• 70+: mood boards, unexpected compositions

--weird — optional taste skew. Low values (50–150) for subtle uniqueness; high values for experimental work.

Don't tune all three every prompt. Set defaults for your project (e.g. raw + s100 + chaos 10 for ecommerce) and change one knob at a time.

Reference parameters (v7 essentials)

--sref [url] — style reference. Transfers palette, texture, photographic grade without copying composition.

• --sw 50–100 — subtle style hint
• --sw 200+ — strong look lock

Multiple: --sref url1 url2 or weighted --sref url1::2 url2::1

--cref [url] — character reference. Same face/body across scenes.

• --cw 75–85 — recognizable but natural pose variation
• --cw 100 — maximum face lock

Best cref sources: clean frontal photo, even lighting, no obstructions.

--oref [url] (Omni Reference) — object/subject lock when cref isn't the right fit. --ow 70–100 for strong match.

Workflow: generate one hero image → use as sref for project look, cref for recurring character → text prompts describe only scene changes.

See our --cref guide on the blog for character-specific depth.

Multi-prompt weights (::)

Separate concepts with :: and assign importance:

ancient temple::3 misty forest::2 moonlight::1

Temple dominates composition; mist secondary; moonlight accent.

Use when: blending two strong ideas without writing a long sentence. Avoid on simple single-subject shots — adds noise.

Negative space: --no

--no text, watermark, logo, extra fingers, blurry

Build a personal --no baseline for your niche (product: "no hands, no props"; portraits: "no distortion").

v7 still prefers positive description over long negative lists. Say "clean white seamless background" instead of only "no clutter."

Aspect ratio is composition

--ar changes framing, not just crop:

• --ar 16:9 — cinematic landscape, environmental storytelling
• --ar 4:5 / 3:4 — portrait photography, social vertical
• --ar 1:1 — product, icons, album art
• --ar 21:9 — ultra-wide establishing shots

Pick ratio before generating — changing later wastes iterations.

Draft Mode: iterate cheap, render once

Draft Mode (~10× faster, lower cost) is for composition and prompt testing. Standard render for final delivery.

Workflow:

1. Draft + short prompt → find composition
2. Refine wording once layout works
3. Add sref/cref if needed
4. Final generation without draft (or upscale path you prefer)

Don't polish prose in draft mode for hours — polish structure, then commit.

What to stop doing in 2026

Keyword stuffing — "8k, masterpiece, trending on ArtStation, ultra detailed, hyperrealistic" often degrades v7. Replace with specific camera, light, and material words.

Copying v4 tag lists — community tag banks from 2022 ignore v7's language model.

Ignoring --v 7 — be explicit on version when comparing or migrating old prompts.

Fighting raw with stylize 800 — pick a lane.

Describing the same character every time — use `--cref` instead of re-pasting face prose.

Skipping image references — if you have a mood board photo, `--sref` beats ten adjectives.

Category templates (copy and edit)

Product (e-commerce)

Ceramic coffee mug matte white glaze on warm oak surface, soft 45-degree studio light, subtle shadow, catalog photography --ar 1:1 --style raw --stylize 120 --chaos 5 --v 7

Environmental portrait

Man waiting at tram stop, overcast European city, candid street photography, 35mm film grain, natural colors --ar 3:2 --style raw --stylize 80 --v 7

Concept art (stylized)

Floating islands connected by rope bridges, sunset god rays, painterly fantasy concept art, atmospheric perspective --ar 16:9 --stylize 400 --chaos 25 --v 7

Brand consistency project

[Scene description] --sref [brand-mood-url] --sw 80 --style raw --ar 16:9 --v 7

From rough idea to finished prompt

1. Write one sentence: who/what, where, light, medium.
2. Add camera or art medium if photographic or editorial.
3. Append parameters: --ar → --style raw (if photo) → --stylize → --chaos → --v 7.
4. Generate Draft → adjust ONE layer (light OR lens OR palette).
5. Add --sref / --cref when repeating.

Shortcut: upload a reference to PromptMake /image, target Midjourney v7, mode Recreate or Variation — get a v7-shaped prompt, then edit parameters. 3 free generations/day on /image.

Reverse-engineering workflow: see our guide on extracting prompts from reference photos.

Troubleshooting

Looks like a painting when I wanted a photo — add `--style raw`, lower `--stylize`, name a real lens.

Ignores my composition — subject later in sentence? Move it first. Chaos too high?

Same face drift across shots — `--cref` with `--cw 85`, cleaner reference image.

Style inconsistent across campaign — `--sref` from approved hero frame, `--sw 80–120`.

Too boring — raise stylize slightly OR add one strong mood adjective, not ten generic ones.

Weird hands/text — `--no text, extra fingers`; for text in scene use Ideogram or post-process.

Midjourney vs doing it in ChatGPT Image

Midjourney rewards parameter control, grid iteration, sref/cref. ChatGPT Image (GPT-5.5 ecosystem) rewards conversational scene description with less parameter syntax.

Use Midjourney when: aesthetic exploration, style-locked campaigns, community grid workflow.

Use ChatGPT Image when: quick integrated text+image, less parameter learning curve.

PromptMake outputs dialect per target — pick Midjourney v7 before paste.

Frequently asked questions

v7 or v6.1 in 2026?

Default to v7 for new work. Migrate old prompts explicitly with --v 7 and retune stylize — don't expect 1:1 parity.

Is --style raw the same as --raw?

Use --style raw in current docs/community. Older shorthand --raw appears in some guides — verify against Midjourney's parameter list for your account tier.

Best stylize for portraits?

With raw: 50–120. Without raw: 150–250 if you want flattering MJ polish.

How long should prompts be?

One to three sentences plus parameters. Longer ≠ better on v7.

Niji vs standard v7?

Niji for anime/illustration (--style expressive, --style cute). Standard v7 for general work. PromptMake has Niji 6 target if needed.

Can PromptMake replace learning MJ syntax?

It accelerates structure from references or rough ideas. You still need parameter literacy for iteration and campaigns.

Related articles

Reverse-engineer AI image prompt from any photo — reference → MJ prompt workflow.

Midjourney v7 guide (legacy) — deep parameter reference for --sref, --cref, Draft Mode.

Midjourney --cref guide — character consistency details.

Midjourney vs Flux (legacy) — when to leave MJ syntax for natural language.

Bottom line

Midjourney prompts that work in 2026: clear natural-language scene direction, subject first, parameters last, --style raw for anything photographic, stylize/chaos tuned intentionally, references for consistency, Draft for exploration.

Drop the masterpiece keyword stacks. Write like you're briefing a photographer or art director — then let v7 do what it's good at.