The Parameter Deck

Aspect ratio locks the frame. Common values are 1:1 for square, 3:2 or 4:5 for print, 16:9 for cinema, and 9:16 for phone stories. It is the one control that changes what fits in the shot, not just how it looks.

The same prompt in 1:1 versus 16:9 will compose entirely differently — subjects reframe, foregrounds crop out, backgrounds open up.

Decide the frame first. Pick 3:2 or 16:9 for cinematic scenes, 4:5 or 9:16 for social, 1:1 for isolated product shots. Never change it mid-refine, or your composition resets.

Every image model has versions. Newer ones understand longer prompts and more nuance but produce a different aesthetic. Niji is a separate model tuned for anime, manga and illustration.

Switching versions changes the whole look of your prompt, not just quality. Old prompts often need rewording on new versions.

Stay on one version through a project. If you're doing anime or illustration, use Niji. For everything else, use the latest general version unless you have a reason not to.

Stylize controls how far the model editorialises your prompt. Low values follow your words literally. High values add lighting, atmosphere and painterly touches the model thinks look good.

At 50 the image reads like a technical brief. At 750 it reads like the model illustrated your brief. Default sits around 100.

For product, spec and reference work, keep it under 100. For editorial, mood and concept work, sit between 250 and 500. Anything above 750 is decorative — treat it as a special effect.

Chaos changes how much variety the model puts across the four images in a run. Low chaos gives you four close cousins. High chaos gives you four different children.

At 0 the grid is nearly interchangeable. At 50 you see genuine variety. At 100 you get wildly different interpretations of the same words.

Explore with chaos high, converge with chaos low. Start a project at 25 to see options, then drop to 0 once you know the direction.

Weird pushes the model away from familiar-looking outputs. It doesn't randomise composition like chaos — it warps the aesthetic itself into stranger territory.

Low values look normal. High values look like art-school experiments — odd palettes, unusual materials, uncanny proportions.

Use it for concept exploration and personal work. Keep it at 0 for anything a client will approve or a factory will build.

Style Raw tells the model to skip its default beautification pass. Colours calm down, lighting gets less dramatic, compositions read more like documentation.

The same prompt with and without raw looks like two different photographers took it — one editorial, one technical.

Turn it on for spec work, materials studies, product shots and anything where accuracy matters more than mood. Turn it off for editorial and hero images.

Style Reference takes an image URL or code and pulls its palette, texture and lighting into your prompt. Style Weight controls how strongly it comes through.

Your subject stays yours, but the mood matches the reference. At weight 100 it's a suggestion, at 500 it's a rewrite.

Reference codes (--sref random or a saved number) are cheaper to iterate on than image URLs. Pick a weight that lets your subject lead — usually 100 to 300.

Character Reference locks a person's likeness across a series. Character Weight controls how much of the reference you carry — just the face, or the whole look.

At cw 0 you get the face and nothing else. At cw 100 you get face, hair, clothing and props from the reference.

Use cw 0 for character sheets in different outfits. Use cw 100 when you need the same person, same outfit, in a new scene.

Negative prompts tell the model to avoid specific things — colours, objects, styles, features. It works by biasing away, not by hard-banning.

Small lists (two or three items) work reliably. Long lists dilute each item and often do nothing.

Only list things that keep appearing uninvited. Don't preemptively list everything you dislike — you'll flatten the image.

Image Weight decides who wins when your uploaded image and your prompt disagree. Low values let the text lead. High values let the image lead.

At iw 0.5 the image is a hint. At iw 2 the image is the brief and the text tunes it.

Start at 1 and adjust. If the model is ignoring your reference, go up. If it's copying it too literally, go down.

Quality is a render-time multiplier. Lower values are faster and cheaper but less detailed. Higher values take longer and cost more credits, in exchange for finer texture.

The composition doesn't change with quality — only how carefully it's rendered. Don't expect a different image, expect a more finished one.

Explore at q 0.5. Finalise at q 1. Only push to q 2 for hero shots where the extra detail is going somewhere it matters.

Every major control surface — the Director Console and the Midjourney parameter deck — has a GRID | RACK toggle. GRID is the labelled tile layout. RACK renders the same controls as studio hardware: rotary knobs for continuous values, XY pads for paired axes, LED steppers for discrete choices and numeric knobs for exact figures. It is a different instrument for the same values, not a different engine.

Switching views never changes your selections. A value set in RACK reads back identically in GRID, and the prompt output is byte-for-byte the same either way. The toggle is remembered per surface.

Use GRID when you're reading and comparing options, and RACK when you already know the axis you want and just want to feel your way along it. Rack is faster for tuning, grid is faster for choosing.

Rotary knobs are dragged vertically — up increases, down decreases — and double-click resets to the default. XY pads set two related axes at once from a single position, with the current coordinate printed under the pad. LED steppers cycle discrete options, and numeric knobs accept typed values when you need an exact figure. Every control is keyboard-operable: focus it and use the arrow keys for single steps, Page Up/Down for coarse jumps, Home to reset.

Each control shows its live value as text as well as position, so nothing depends on reading a needle. Values are clamped to the range the target engine actually accepts.

Coarse-set with the mouse, then fine-tune with arrow keys. Trying to hit an exact number by dragging wastes more time than tabbing to it and typing.

The SEQUENCE // strip along the rack is a row of LEDs, one per control group. A lit LED means the group is live; clicking it mutes the group. A muted group is excluded from the compiled prompt and skipped entirely by randomization and Auto-Suggest, while keeping the values you set inside it.

Muting is non-destructive. Unmuting a group brings back exactly the values it held before, so you can A/B whether a whole axis — lighting, environment, surface — is helping or fighting the shot.

When a prompt is over-specified and results feel muddy, mute groups one at a time rather than deleting selections. You'll find the group causing the noise in two or three passes, with nothing lost.