Chaos, Weird and Stylize: Midjourney's Creativity Dials

TL;DR: Midjourney stylize runs 0 to 1000 with a default of 100, chaos runs 0 to 100 with a default of 0, and weird runs up to 3000 from a default of 0. All three are ticked for V8.1 and V8.2. Chaos varies images against each other, stylize varies toward Midjourney's aesthetic, weird varies toward the unusual. Verified September 2, 2026.

What do --chaos, --weird and --stylize actually change?

Three different axes, which is why stacking them blindly produces mush. Each one moves a distinct thing, and Midjourney describes each in its own article without ever putting them side by side.

Chaos varies your four images against each other. Midjourney's Chaos article is explicit that it operates on the grid: "Chaos (also called Variety on midjourney.com) lets you add more variety to the image results you get from each prompt." At the default, "By default chaos is set to 0, and Midjourney creates four images based on your prompt." Turn it up and the four diverge. The cost is stated in the same article: "You can set the chaos value anywhere between 0 and 100, but keep in mind that higher values mean the images can be quite different and may not stick as closely to your prompt, giving you unpredictable results."

Stylize varies each image toward Midjourney's house aesthetic. The Stylize article frames it as a trade rather than a quality knob: "stylize lets you choose if you want your image to be more literal (based on your prompt) or more creative and artistic." Low stylize buys prompt adherence. High stylize buys Midjourney's taste, and spends your prompt's specificity to pay for it.

Weird varies toward the unusual. Midjourney calls it "an experimental feature" and warns that what counts as weird "may change over time". That sentence is doing more work than it looks like. It means a --weird value copied from a tutorial is not promised to render the same way today.

On the web interface all three sit under one settings group Midjourney labels Aesthetics, described as: "Play with stylization to influence how artistically your images are created. Adjust weirdness to add unique quirks. Control variety for more diverse outputs."

What are the documented ranges and defaults for each dial?

Six numbers, and all six are published. Here they are with the article each one comes from, rather than the usual undated table.

Two footnotes on that table, both from Midjourney.

The stylize range is confirmed twice more outside its own article. Both the Personalization and Moodboards pages state that "The stylize parameter accepts values from 0 to 1000." That is unusual in this documentation set, and it makes the stylize figure the most solid of the three.

The weird range is stated as a ceiling, not a span. The Weird article's exact wording is "By default, weird is set to 0, but you can add values up to 3000." No page states a minimum, and no page states what happens if you send 5000.

Which Midjourney versions support each dial?

All three, on everything current. This matters more than it sounds, because several neighbouring Midjourney parameters are stranded on older models.

The authority is the Feature Compatibility chart inside the Version article, which was re-saved on 2026-09-01. Read carefully, because the ticks and no-symbols in that chart are image icons rather than characters. Extract them as text and the table looks empty.

The Version article states plainly: "The current default Midjourney version is V8.2." Each of the three parameter articles also carries its own compatibility banner reading that the feature is supported in V8.2, so chart and article agree here. That is not always true in this documentation set, and our quality and speed reference covers a case where they flatly do not.

Weird carries one extra version statement in its own body: "Weird is compatible with Midjourney versions 5 and later." The Legacy chart backs this up by leaving the Weird row blank for V4, V3, the test models and Niji 4. If you are switching models mid-project, our version flag reference has the full --v and --niji picture.

Does Midjourney publish how the three dials interact?

No. Not one sentence, anywhere in the 105 articles currently in the help centre.

This is the honest centre of the whole topic, and it is the thing every best-Midjourney-settings post quietly skips. There is no recommended pairing, no statement about whether high chaos amplifies or drowns high weird, and no note on whether stylize and weird compete for the same influence budget. The single place Midjourney puts two of these dials in one prompt is the Transitioning to Web article, and only incidentally: a Discord option set that expands to sitting in a tree --c 20 --ar 2:3 --s 800, offered as a demonstration of option sets rather than of the combination.

What does exist is a short list of documented interactions, all of them involving stylize and none of them involving each other:

• Stylize doubles as a Personalization strength dial. The Personalization article states that "the stylize parameter controls how much of your Personalization is applied to the images", and the Moodboards article says the same for moodboard style. So on a personalized account, --s is no longer only a house-style knob.
• Stylize competes with Omni Reference Weight. The Omni Reference article warns: "If you're using high stylize or --exp values you may want to also use a higher Omni Reference Weight, as these will all compete for influence." Note the scope. The compatibility chart marks Omni Reference unsupported on V8.1 and V8.2, so the one documented competition statement about stylize applies to a feature you cannot use on the current default model.
• Image-only prompts disable two of the three. The Image Prompts article states: "Prompts that only use images and no text are not compatible with the --stylize or --weird parameters." Chaos is not named in that exclusion.
• Text rendering prefers low stylize. The Text Generation article's remedy for wrong lettering is to "consider using Raw or lowering the Stylize value".

Everything else you have read about combining these dials, including everything below, is somebody's observation. Ours included.

How do you test one dial without fooling yourself?

Change one dial, hold everything else, and compare batches rather than single images. That is the whole method, and the discipline is in the caveats rather than the steps.

1. Write one control prompt and never edit its words. Every difference you see should be attributable to the parameter block.
2. Lock a seed. Append --seed 1234 so the starting noise stays constant across runs.
3. Stay out of Turbo. The Seeds article is blunt: "Don't use Turbo mode when using seeds." On V8.1 and V8.2 this is academic, since the compatibility chart marks Turbo unsupported there anyway.
4. Move one value at a time, in visible steps rather than increments of five.
5. Judge the grid, not the pick. Chaos in particular changes how much the four images differ, which is invisible if you only look at your favourite one.

What does a --chaos ladder look like in practice?

Here is a ladder, and here is the disclaimer that belongs with it: the value bands below are our own observation from running these prompts, not Midjourney documentation. Midjourney publishes the range and the default and stops. Anyone presenting a chaos value table as official is presenting their taste as a spec.

The upper end is exactly what the documentation warns about. Higher values "may not stick as closely to your prompt". That is a feature when you are hunting and a defect when you are delivering.

What stylize values should you actually use?

Start from Midjourney's own presets rather than a round number. The Discord settings panel ships four, and the Creation Settings article publishes the values behind them: "Stylize Low = 50, Stylize Med = 100, Stylize High = 250, Stylize Very High = 750". Med is the default.

That is the only stylize ladder Midjourney publishes, which makes it the right skeleton. The commentary below is ours.

That last row is not our invention. The March 17, 2026 V8 Alpha changelog encouraged testers to use personalization and push stylize to 1000, in the specific context of leaning on Midjourney's stylization systems. It is a vendor recommendation for a particular workflow, not a general default.

For photographic realism, reach for --raw before you reach for --s 0. The Raw article says it produces "more realistic, photo-like images" with simple prompts, and that "Raw is compatible with Midjourney versions 5.1 and later". Our --style raw breakdown covers what changed when the flag was renamed.

How far up the --weird scale is usable?

Further than you would guess, and less linearly than the other two. Once more: the bands below are observation. Midjourney publishes a ceiling of 3000 and a default of 0, and nothing in between.

Because weird is documented as experimental and explicitly not fully seed-compatible, treat any weird value as a batch setting rather than a repeatable one. Save the images, not the number.

Which dial should you reach for first?

Pick by what is wrong with your current grid, not by taste.

If your four images are too similar, that is chaos. If they look like Midjourney instead of like your prompt, that is stylize, and you want it lower. If they are competent and boring, that is weird, and it is the only one of the three that adds something rather than redistributing it.

28 copy-paste Midjourney prompts by intent

One control subject runs through the first three groups so you can read the ladders against each other. Swap the subject, keep the parameter block.

The control. Run this first so every comparison has a baseline.

a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210

Safe and commercial. Low variance, high prompt adherence, predictable delivery.

# 1. Client-safe default. Nothing but the model's balance point.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 0 --s 100

# 2. Literal pass. Stylize at Midjourney's Low preset value.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 0 --s 50

# 3. Maximum prompt adherence. Stylize floor.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 0 --s 0

# 4. Four near-identical options for a layout that is already approved.
packaging mockup of a matte black coffee tin on a concrete slab, soft key light --ar 1:1 --v 8.2 --seed 88 --c 0 --s 50 --raw

# 5. Editorial portrait, house style dialled back so wardrobe notes survive.
half-length portrait of a chef in a white jacket, steel kitchen behind, window light --ar 4:5 --v 8.2 --s 50 --c 0

# 6. Architectural elevation. Literal geometry beats atmosphere here.
straight-on elevation of a timber-clad cabin, overcast noon light, no people --ar 16:9 --v 8.2 --s 0 --raw

# 7. Repeatable series shot. Same seed, same dials, swap only the subject noun.
a hand-thrown ceramic mug on raw linen, single window light --ar 1:1 --v 8.2 --seed 4210 --c 0 --s 100

Exploratory. You know the subject, not the shot.

# 8. First exploration step. Small spread, prompt still governs.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 10 --s 100

# 9. Composition hunt. Four genuinely different framings.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 25 --s 100

# 10. Direction hunt. Expect the brief to be interpreted, not followed.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 50 --s 100

# 11. Chaos ceiling. Four ideas, weak adherence. Throw away three.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 100 --s 100

# 12. Spread plus house style. Both dials up, for mood boards.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --c 40 --s 250

# 13. Stylize High, chaos off. Isolates the aesthetic axis.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 0 --s 250

# 14. Stylize Very High. Prompt details start dropping out here.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --seed 4210 --c 0 --s 750

# 15. Stylize ceiling, the value the V8 alpha post recommended for personalized accounts.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --c 0 --s 1000

Deliberately strange. Weird is the only dial here that adds rather than redistributes.

# 16. Weird floor. Enough for a quirk, safe for commercial work.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --w 100 --s 100

# 17. The band most people actually want. Odd but coherent.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --w 750 --s 100

# 18. Surreal. Subject survives, physics does not.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --w 1500 --s 100

# 19. Weird ceiling. Mostly unusable, occasionally the best frame you get.
a lone lighthouse on a basalt headland at dusk, low sea mist, long exposure --ar 3:2 --v 8.2 --w 3000 --s 100

# 20. Weird with low stylize. Strange composition, literal rendering.
a Victorian greenhouse growing television sets, damp glass, morning light --ar 3:2 --v 8.2 --w 1000 --s 0 --raw

# 21. Weird with high stylize. Both interpretive layers on at once.
a Victorian greenhouse growing television sets, damp glass, morning light --ar 3:2 --v 8.2 --w 1000 --s 750

# 22. All three up. Undocumented territory. Read the grid, not one image.
a Victorian greenhouse growing television sets, damp glass, morning light --ar 3:2 --v 8.2 --c 50 --w 1500 --s 500

Style-neutral, for photographic realism.

# 23. Raw plus stylize floor. The most literal combination available.
documentary photograph of a fishing boat unloading crates at 6am, harbour light --ar 3:2 --v 8.2 --raw --s 0 --c 0

# 24. Raw at the default. Realism without flattening the light.
documentary photograph of a fishing boat unloading crates at 6am, harbour light --ar 3:2 --v 8.2 --raw --s 100 --c 0

# 25. Realism with a little grid spread, for choosing a frame.
documentary photograph of a fishing boat unloading crates at 6am, harbour light --ar 3:2 --v 8.2 --raw --s 50 --c 15

# 26. Product realism. Raw carries the material, stylize stays out of it.
studio photograph of a brushed steel watch case on grey seamless paper, softbox left --ar 1:1 --v 8.2 --raw --s 0

# 27. Text in image. Midjourney's own remedy is Raw or a lower stylize value.
a hand-painted shopfront sign reading "HARBOUR BOOKS", weathered green paint --ar 3:2 --v 8.2 --raw --s 50

# 28. The A/B harness. Run these two, change nothing else, compare the grids.
a lone lighthouse on a basalt headland at dusk, low sea mist --ar 3:2 --v 8.2 --seed 4210 --c 0 --s 100 --w 0

Where do these three dials stop working?

Three places, and one of them is a contradiction inside Midjourney's own documentation.

On variations and remix, probably. The Remix article states: "Only parameters that normally influence variations will work while using Remix, including: aspect ratio, no, stop, and tile." None of our three dials is on that list. The Legacy Features chart agrees, ticking Chaos, Stylize and Weird under Affects Initial Images and leaving Affects Variations and Remix blank.

Except in Vary Region, where the same documentation says the opposite. The Vary Region article states: "The following parameters are compatible with Vary Region when editing your prompt: chaos, generation speeds, image weight, no, stylize, style, version, video, and weird." The Remix article names Vary Region as one of the tools Remix covers, so the two pages disagree about the same operation. Both are live. We are reporting the conflict rather than picking a winner.

On image-only prompts. Two of the three are ruled out entirely: "Prompts that only use images and no text are not compatible with the --stylize or --weird parameters."

What Midjourney does not publish about chaos, weird and stylize

Writing the gaps down is more useful than filling them, because every one of these is currently answered with confidence somewhere on the first page of results.

• Any interaction between the three. No combined example, no recommended pairing, no statement of precedence anywhere in the help centre.
• Any recommended value for any of them. The only published ladder is the Discord stylize preset set. There is no official "good" chaos or weird value.
• What happens out of range. No page says whether --w 5000 errors, clamps to 3000, or is ignored.
• A lower bound for weird. The article gives a ceiling and a default, never a floor.
• What "not fully compatible with seeds" means in practice. The caveat is stated and never explained.
• Any GPU cost for these parameters. Midjourney publishes per-job GPU minutes and attaches none of them to chaos, stylize, weird or raw.
• Niji 6 and Niji 7 behaviour. Neither model appears in either compatibility chart, so per-version support for these three on Niji is simply unstated.
• The alpha site's House style control, which appears in the August 20, 2026 changelog and in zero help-centre articles.

How to check all of this yourself

Midjourney re-saved 18 of its 105 help articles in the eight days before this post went up, the Version chart, the Parameter List and Legacy Features among them, so verification beats memorisation here.

Read the compatibility chart from raw HTML. The ticks and no-symbols in the Version article are <img> icons. Any tool that strips a page to text hands you an empty table, and an empty cell reads exactly like an unsupported feature.

Do not read chart silence as a verdict. Some features have no row in that chart at all, and absence means the article's own banner is the only support statement, not that the feature is unsupported.

Fetch the docs through the Help Center JSON API. docs.midjourney.com returns 403 to most automated requests, which is not a dead link. The updated_at timestamp in that JSON is the only reliable freshness signal, and it is why this post dates every article it quotes.

Check the settings panel before you trust a prompt. Stylize, weirdness and variety all have persistent defaults in the Aesthetics group, and a saved default silently applies to every prompt that does not override it. Most reports of a prompt that will not reproduce come down to that.

Once a parameter block earns its place, save it as a reusable template with the subject as a variable rather than retyping it and drifting a digit. That is the same habit that makes an image prompt template library worth keeping.

Where we fit, since it should be said plainly: Prompt Architects does not generate images. It generates the prompt, and you paste that into Midjourney. Image prompt generation is a paid feature on our side, available from Pro upward per the comparison table on our pricing page, while the Free plan covers 5 prompt enhancements per day, forever, per our FAQ. Saving a parameter block to an image prompt library needs Advanced.