2.2--stylize, --weird, --chaos, --seed
The essentials in 30 seconds
- --stylize arbitrates: low = faithful to the prompt, high = aesthetic but free — no ideal value, an arbitration for each intent.
- --chaos explores, --weird surprises: diversity of the 4 options and dose of strangeness — explore wide, then converge.
- --seed freezes randomness: same seed + same prompt + same settings = reproducible — one change at a time.
Running mini-case — Sonia isolates the effect of one word
Sonia likes one direction of her Algiers café poster but hesitates on the light: she recovers the image seed, freezes everything, then changes ONLY one word (morning light versus evening light) and reruns. She finally sees the real effect of the word instead of blaming randomness. Earlier, she had explored wide (high --chaos) then lowered --stylize to stay faithful to her composition. In 2.3 she will keep this method to dose --iw; in 2.4 she will freeze her series parameters; in 2.5 she will log version + seed in her library.
Sequel to thread 2.1: Sonia's poster moves from suffered randomness to controlled experiment — and continues in 2.3, 2.4 and 2.5.
1. Stylize and chaos: arbitration and exploration
--stylize (--s) sets the intensity of Midjourney's own 'taste': a low value = more faithful to the prompt (literal render, sometimes flatter), a high value = more aesthetic but freer (elements reinterpreted). It is the central arbitration: detailed and demanding prompt → lower; short and suggestive prompt → raise. --chaos (--c) controls the diversity of the four options: 0 = very similar, high value = very different directions — precious in the exploration phase, like Sonia sweeping wide before choosing her direction.
2. Weird and seed: surprise then reproduce
--weird (--w) introduces strangeness to escape expected renders; at high values results become frankly experimental. --seed fixes the random seed: same seed + same prompt + same version + same parameters = reproducible or very close result. It is the methodical-work tool: Sonia freezes the seed of an image she likes, changes a single word, and precisely measures its effect — without a seed, impossible to know whether a change comes from her or from randomness.
3. The method: explore then refine
Two phases: explore (high --chaos and --weird, free seed, re-rolls) to discover directions, then refine (low --chaos, frozen --seed, one change at a time) to converge. To recover a seed: on Discord, react with the envelope emoji; on the web, see the image details in the gallery. And parameters change with versions: Sonia retests her settings at each major version and logs everything (prompt, seed, version) in her library.
Did you know?
The seed turns opinion into measurement.
🔎 Concrete example — 🔎 Worked example — Isolating the effect of one word
Like a controlled experiment, on Sonia's poster:
café poster, steaming cup, morning light --ar 4:5 --seed → note the render
same prompt, 'evening light' instead, same --seed → compare
→ you precisely measure the effect of the word 'evening'. Reading: only the changed word explains the difference, not randomness.
Lesson: without a frozen seed, impossible to know whether a change comes from you or from the draw.
Section vocabulary
- --stylize
- Intensity of Midjourney's own style: low = faithful to the prompt, high = more aesthetic but freer.
Simple example: Sonia lowers --stylize to keep her exact composition.
What it is not: a format or diversity setting.
Reflex: demanding = low stylize; mood = high stylize.
- --chaos / --weird
- --chaos sets diversity across the 4 options; --weird adds a dose of strangeness to escape expected outputs.
Simple example: Sonia raises --chaos to sweep wide, --weird for an unexpected direction.
What it is not: a way to freeze or reproduce an image.
Reflex: explore = high chaos/weird.
- --seed
- Random seed; fixing it makes the result reproducible and lets you isolate the effect of a change.
Simple example: Sonia freezes her poster seed then changes a single word.
What it is not: a guarantee of exact duplication across different versions.
Reflex: refine = frozen seed, one change.
- Exploration versus refinement
- Two phases: explore (high chaos/weird, free seed) then converge (low chaos, frozen seed).
Simple example: Sonia explores 4 re-rolls then refines the selected direction.
What it is not: changing five elements at once while hoping to conclude.
Reflex: first wide, then one by one.
In practice — Hands on — Steering the render with parameters
- Take a working prompt and generate it with low --stylize, then high: watch the fidelity/aesthetics arbitration.
- Rerun with high --chaos to see the diversity of options (exploration phase).
- Recover the seed of an image you like, then rerun the same prompt with that --seed.
- With the seed frozen, change ONE word of the prompt and compare: you measure the real effect of that word.
Ready-to-use script — explaining the seed method to a colleague (30s)
'For Sonia's poster, we stopped changing everything at once: we froze the seed of the direction we liked, then changed a single word — the light. Then we saw the real effect of the word, not randomness. Before that, we had explored wide with high chaos. Explore, freeze, isolate: that is the controlled experiment of prompting.'
Say while showing both images at constant seed, one word apart.
Copy-paste template — explore then refine
[exploration] [subject], [context], [light] --ar 4:5 --chaos high --stylize high
[comparison] same prompt --ar 4:5 --stylize low → compare fidelity / aesthetics
[freeze] same prompt --ar 4:5 --seed → reproduce the direction
[isolate] same prompt + ONE word changed --ar 4:5 --seed → measure the effect Poster faithful to the brief
Imposed café composition: Sonia lowers --stylize so the prompt is followed literally.
Mood illustration
Atmosphere visual with no constraint: Sonia raises --stylize and lets Midjourney beautify.
Unexpected direction
Series too tame: Sonia adds --weird to escape expected renders and find a fresh direction.
Prompt library
Each try logged (prompt + seed + version): Sonia reproduces her series even after a new version.
✓ Best practices
- Lower --stylize for fidelity, raise it for aesthetics — test low versus high on YOUR prompt.
- Explore with high chaos/weird and free seed, refine with low chaos and frozen seed.
- Check the up-to-date documentation: names, ranges and behaviors change with versions.
✗ Common pitfalls
- Changing several elements at once: the effect becomes unisolable.
- Looking for a universally 'ideal' stylize value instead of arbitrating for each intent.
- Forgetting that the seed only reproduces if everything else is identical (prompt, version, parameters).
Key takeaways
- --stylize arbitrates between prompt fidelity (low) and render beauty (high).
- --chaos sets the diversity of the 4 options; --weird adds strangeness to leave beaten tracks.
- --seed freezes randomness: essential to test one change at a time.
- Method: explore (high chaos/weird, free seed) then refine (low chaos, frozen seed). Parameters change with versions.
48h challenge — apply tomorrow
On a real prompt: try low versus high --stylize, explore with high --chaos, freeze the --seed of one direction, then change a single word and log prompt + seed + version in your library.
- I can arbitrate fidelity/aesthetics with --stylize for my intent.
- I can explore (--chaos/--weird) then converge (frozen --seed).
- I log prompt, seed and version to reproduce.
Frequently asked questions
How do I recover and reuse the seed of an image I like?
On Discord, react with the envelope emoji: the bot sends the seed privately, to copy with --seed. On the web, see the image details in the gallery. The seed only reproduces if everything else is identical — which is what lets you isolate the effect of a single change.
Which --stylize value should I choose? Is there an ideal setting?
No: it is an arbitration for each goal. Low = literal and faithful (exact composition, brand constraint); high = beautiful but free (mood illustration). Test the same prompt at two distant values, then adjust — and redo the test at each major version.
What is the --seed parameter for?