QWE AI Academy Three hours. That’s how long I spent trying to remove a backpack from a Midjourney image using --no backpack. V6 refused. V5.2 nailed it first try – then added a survivalist I never asked for.
That’s when I realized most parameter tutorials are teaching you the wrong mental model.
The Chaos Paradox Nobody Mentions
You’re working on a character design. Perfect seed locked in. Now you want variations, so you add --chaos 50 expecting wild diversity.
Doesn’t work that way. When you lock a seed and add chaos, Andrei Kovalev’s chaos study found something weird: “chaos seems to add order to the variations.” The four images become systematically different – pushed away from center in distinct directions – not randomly scattered.
Think of it this way: no chaos with a seed? Four similar photos of the same person. High chaos with that seed? Same person in four deliberately different poses and lighting setups. Controlled variation. Not random mess.
Without a seed? Chaos just scatters results across the entire solution space. Hit --chaos 100 and you’ll get four images that barely relate to your prompt.
Which raises a question: if adding constraints makes chaos more useful, what does that say about how we’re supposed to use randomness?
How Chaos, Weird, and Stylize Differ
Every tutorial lists all three. Few explain when to use which.
| Parameter | Controls What | Range | Default |
|---|---|---|---|
--chaos or --c |
Diversity between the 4 grid images | 0-100 | 0 |
--weird or --w |
Unusual qualities within each image | 0-3000 | 0 |
--stylize or --s |
How much Midjourney’s aesthetic is applied | 0-1000 | 100 |
The official Midjourney docs say chaos “lets you add more variety to the image results.” Here’s the part they bury: higher values mean images “may not stick as closely to your prompt, giving you unpredictable results.”
Weird is different – tells the AI to take creative risks on content itself. Face might have three eyes. Building might melt. But all four grid images will be consistently weird in different ways.
Combine
--chaos 30 --weird 500for exploration without losing your prompt. Chaos spreads concepts; weird bends execution. Together: diverse unusual results that still reference what you asked for.
Setting Up Parameters (And the Syntax Traps)
Parameters go at the end. Format: /imagine prompt: [description] --parameter value
/imagine prompt: cyberpunk street market, neon signs, rain --chaos 40 --stylize 200
Common mistakes:
- No space before dashes:
poppies--chaos 50fails - Extra spaces between dashes:
-- chaos 50breaks - Punctuation after:
--chaos 50,silently ignored - Parameters before prompt: Midjourney skips them
Web interface (midjourney.com) calls chaos “Variety” in settings. But you still type --chaos or --c in Discord. Shorthand works: --c, --w, --s.
When Chaos Hurts Your Results
Tested with architecture prompt: “isometric city block, colorful pipes, trees.”
--chaos 0: Four similar views. All clearly city blocks. All following prompt.
--chaos 25: Color and angle variation. Same concept.
--chaos 75: One image, no buildings. Another, no pipes. Third lost isometric view.
--chaos 100: Four images sharing maybe two elements from original prompt. One was just rainbow industrial machinery in glass.
Chaos past 50? AI starts ignoring chunks of your prompt for “artistic freedom.” By 75, you’re giving a suggestion, not a specification.
Use chaos above 50 only when brainstorming and you genuinely want surprises. Client work or specific needs? Stay under 30.
The Weird Version Problem
No beginner tutorial mentions this: --weird behaves completely differently depending on which Midjourney version you’re using.
“V5.2 might just be the front-runner for the title of ‘weirdest Midjourney model.'” That’s from the Midlibrary V6 review. V6 brought “a greater sense of consistency” to weird outputs – polite way of saying V6 weird is less weird.
V7 (default since June 17, 2025) and V8 Alpha (launched March 17, 2026 on alpha.midjourney.com) continue this. Want truly bizarre results? Switch back to V5.2 with --v 5.2 --weird 2000.
Docs warn “what’s ‘weird’ may change over time” and that weird is “experimental.” Translation: they’re still tuning it. No guarantee your --weird 1500 prompt produces the same vibe six months from now.
One more: weird isn’t fully compatible with seeds. Same seed + weird value = different results. Can’t reproduce reliably.
The –no Parameter Trap
You’d think --no modern clothing would remove modern clothing.
Nope.
Midjourney’s moderation system reads every word in the –no parameter independently (per the official –no docs). So --no modern clothing becomes TWO commands: “no modern” + “no clothing.”
Result? Historical figures in loincloths. Or all fabric removed.
Better: --no jeans, sneakers, t-shirt. Specific items, not compound phrases.
Even then, –no is unpredictable. Midlibrary V6 study tested removing backpacks and cars from city scenes. V6 couldn’t remove the backpack. V5.2 succeeded but added random objects.
Use –no for broad elements (no trees, no people, no text) rather than precise exclusions. Need surgical control? Use inpainting in the Editor.
Tile: The Pattern Parameter
Less common but useful: --tile makes smooth repeating patterns for wallpaper, fabric, game textures.
Key detail from the official tile docs: “Upscaling your tile images will often break the smooth repeating pattern, so we do not recommend upscaling these images.”
Need high-res tiled patterns? Generate at highest base resolution (--hd in V8 Alpha), then tile externally in Photoshop or a pattern tool. Don’t rely on Midjourney’s upscaler.
Combining Parameters: What Works, What Breaks
Chaos and stylize play well together. Kovalev’s study: “–chaos is no less efficient way to experiment and enrich your generations than –stylize, and they do work well together.” At --chaos 50 --stylize 500, diverse results that all maintain high aesthetic quality.
Chaos and weird stack – but carefully. --chaos 100 --weird 3000 gives you four images with almost nothing to do with your prompt. Fever dreams. Fun for art projects, useless for commercial work.
Weird and seed don’t combine reliably. Docs confirm weird “is not fully compatible with seeds.” Don’t expect reproducible weird images.
Client concept exploration setup I use: --chaos 20 --stylize 150 --seed [saved seed]. Enough variety to show options, enough control to stay on-brand.
What Nobody’s Measuring
Tutorials show you grids at chaos 0, 25, 50, 100. But how does chaos interact with prompt complexity?
Simple prompt (“red apple”): Chaos has minimal impact until 60+. Concept too constrained.
Complex prompt (“Victorian street scene, gas lamps, fog, horse-drawn carriage, cobblestones, top hats”): Chaos at 30 already starts dropping elements. At 50? Might lose carriage or fog entirely.
More concepts in your prompt = lower safe chaos threshold. Prompts with 5+ elements? Stay under --chaos 20 if you want them all to appear.
Honest Limitations
Chaos doesn’t give you “more creative” images. Gives you more VARIED images, some of which ignore parts of your prompt. Tutorials conflate variety with quality.
Weird has gotten less weird over time as Midjourney prioritizes coherence. Want truly strange outputs? You’re fighting the model’s current training direction.
–no is a suggestion, not a command. AI will try, but multi-word phrases break its parsing. Even single-word exclusions fail randomly.
Tile breaks on upscale. Limits usefulness for print-resolution patterns.
None of these parameters give pixel-level control. Vibes, not CAD tools.
FAQ
What’s the real difference between chaos and weird?
Chaos controls how different the four grid images are from each other. Weird controls how unusual each individual image looks. Different axes entirely.
Why does combining chaos with a seed make images LESS random?
When you lock a seed, you’re fixing the starting noise pattern. Chaos then pushes the four variations away from that center point in systematic directions rather than scattering them randomly. Community testing shows this creates “ordered” variation – same concept, deliberately different executions – instead of true randomness. Turns out it’s more useful for iteration than pure chaos, which is counterintuitive but makes sense once you’ve tested it a few times with the same seed.
Can I use –no to remove specific objects reliably?
No. Multi-word phrases get parsed wrong. Even single words fail – V6 couldn’t remove backpacks in testing. Use –no for broad stuff (trees, people, text). Editor’s inpainting for precision.
Next: Pick one image you’ve already generated. Add --chaos 25 and run it again with the same prompt. Compare what changed – not just composition, but which parts of your prompt got emphasized or dropped. That’ll teach you more than any parameter chart.