You type a prompt, hit enter, and get four images that are almost right. The lighting is flat, one hand has six fingers, and nothing matches the picture in your head. So you reroll, swap a word, reroll again, and twenty minutes later you have burned through a chunk of your fast hours with nothing usable to show for it.
That loop is avoidable. Midjourney rewards a specific working order: decide the frame first, describe in layers instead of sentences, then narrow with references and parameters. Here is the sequence I use, with prompts you can copy, paste, and edit.
Step 1: Decide the frame before you write a single word
Most people write the subject first and think about the aspect ratio never. Flip that order. The –ar flag changes how Midjourney composes everything else in the image, because it has to fill a different shape with a subject. A portrait cramped into a 16:9 frame gets pushed to the side with dead space around it.
- 16:9 for a blog header or a YouTube thumbnail
- 4:5 for Instagram feed posts, which crop most vertical images badly
- 2:3 for posters and book covers
- 1:1 for avatars and product tiles
While you are setting the frame, pick your stylize value. –stylize runs from 0 to 1000 and defaults to 100. Anything under 100 keeps Midjourney literal and obedient. Push it to 400 or 750 and it starts making aesthetic decisions on your behalf, which is great for moody editorial work and terrible when you need a specific product shape. If you are still deciding which plan gives you enough generations to experiment, the breakdown of Midjourney’s prompts, pricing tiers, and roadmap is worth ten minutes before you commit.
Step 2: Build the prompt in three layers
Here is the mistake I made for months. I wrote full sentences, the way you would describe a photo to a friend. Midjourney blends that into one soft blob of meaning and gives you something vaguely related. Fragments weight better because each one registers as its own signal.
Build in this order:
Layer one: the subject
Who or what, doing what, where. Be concrete about age, material, and action. “An elderly fisherman mending a net on a harbour wall” beats “a fisherman” every time.
Layer two: medium and light
This is where image quality actually comes from. Name the light source, not the mood: “overcast North Atlantic light,” “single tungsten lamp from the left,” “harsh midday sun through blinds.” Then name the capture or medium: “shot on 85mm f/1.8,” “gouache on cold-press paper,” “35mm film grain.”
Layer three: the technical finish
Colour direction and parameters go last. “Muted blues and rust” followed by --ar 4:5 --stylize 300 --v 6.1.
Assembled, it looks like this:
an elderly fisherman mending a net on a harbour wall, worn wool sweater, overcast North Atlantic light, shot on 85mm f/1.8, shallow depth of field, muted blues and rust --ar 4:5 --stylize 300 --v 6.1
That structure holds up across styles, which is part of why the tool has held its grip on illustrators while competitors come and go. The reasons behind that staying power get a proper look in this piece on why Midjourney remains the artist’s choice for AI imagery.
Step 3: Lock the look with style and character references
Once a prompt works, you rarely want to gamble on it again. Two flags fix that.
–sref takes an image URL and borrows its palette, texture, and overall treatment. Pair it with –sw (style weight, 0 to 1000) to control how hard it leans. A value around 100 keeps your subject dominant; 700 will repaint almost everything in that reference’s aesthetic.
–cref does the same job for a character’s face, and –cw (character weight, 0 to 100) decides whether it copies just the face or the whole outfit too. Set --cw 0 when you want the same person in different clothes.
Then save your –seed. React to your favourite job in Discord with the envelope emoji and Midjourney DMs you the seed number. Drop that number back into a prompt and you get the same starting noise, which means small prompt edits produce small visual changes instead of an entirely new picture. Seeds plus a style reference are how people build anything that needs to look consistent across fifty images.
Step 4: Iterate with the right tool, not the same tool twice
Rerolling is not iteration, it is gambling. Each button in Midjourney solves a different problem:
- Reroll (the circular arrows): use when the composition is right and you just want another pass at the details. Twice, maximum, then change your words.
- Vary Subtle / Vary Strong: Subtle for a near-miss, Strong when you like the idea but not the execution.
- Vary Region: the fix for a single bad hand, a mangled ear, or a floating object. Select just that area and describe only what belongs there.
- Zoom Out 2x and Pan: when your subject is cropped too tight and you need breathing room around it.
- Remix mode: lets you edit the prompt while generating a variation, so you keep the composition and swap the lighting.
One habit worth building: write down the exact prompt and seed whenever something works. Reconstructing a good result from memory three days later is close to impossible.
Step 5: Fix the failures everyone hits
Some problems are structural, not bad luck.
Hands. Avoid prompts that put hands near the camera or overlapping other objects. Vary Region is faster than rerolling for a single hand. Fingers holding something thin, like a pen or a wire, will fail more often than a relaxed open palm.
Text. Midjourney still produces lettering that looks like a language you have never seen. Generate the image clean, then set the type in Canva, Photoshop, or Figma. You will get better kerning and a correct spelling of your own brand name.
Symmetry. Architecture and faces both drift. If you need a symmetrical building, say so and add --no asymmetry. Expect to Vary Region on the windows anyway.
Glassy eyes at distance. Any face smaller than about a fifth of the frame loses detail. Frame tighter, then Zoom Out once you have a version you like. There is a good rundown of the remaining weak spots in AI image generation in this look at where AI drawing really breaks and how artists work around it.
Step 6: Turn one good result into a set of twenty
Say you are illustrating a small board game and you need twelve consistent character cards. Generate one you love, then freeze everything except the subject line.
Keep this tail identical across every prompt: , painterly fantasy illustration, warm lantern light, muted earth palette --ar 2:3 --sref [URL of your winner] --sw 150 --seed 4471
Now swap only the opening:
a fox archer in autumn woodsa badger blacksmith at a forgea heron apothecary in a cluttered shop
Twelve prompts, one visual language, no drift between cards. The same trick works for a brand’s social graphics, a children’s book, or a series of article headers. If you want a set to keep some of your own hand in it rather than looking uniformly machine-made, the approach in this guide to making AI-assisted art without losing your own style is a useful counterweight.
Build a prompt library, not a prompt graveyard
The real advantage compounds over months. Keep a plain text file, one entry per working prompt, with the full string, the version flag, the seed, and a note on what you were trying to make. After a few weeks you will notice patterns: the stylize range that suits your eye, the lighting phrases that actually shift an image, the subjects that reliably need region edits.
When a new model version lands, your library becomes a test rig. Run five old favourites through it and you will know within minutes whether your workflow survives the upgrade or needs rebuilding. That is a much better position than starting from a blank prompt box every time, hoping this is the roll that finally matches the picture in your head.