How to Get Genuinely Good Images Out of Midjourney V7

1. Write sentences, not keyword soup

This is the single biggest shift from V6, and most tutorials still haven’t caught up. In V6, keyword stacking was a defensible strategy. Throw enough descriptors at the model and something would stick. V7’s improved language understanding changed the calculus. Descriptive sentences like “a woman reading in a sun-lit café corner, late afternoon light angling through tall windows” produce better results than “woman, reading, café, sunlight, warm, cozy.” Write the scene the way you’d describe it to a set photographer, not as a tag cloud.

Stop typing “cinematic, moody, atmospheric, dramatic lighting, ultra detailed.” That sentence tells V7 nothing. “A detective standing at a rain-streaked window, single desk lamp behind him casting his shadow on the wall” tells it everything. V6 and V7 understand natural language more accurately, so over-stuffing prompts with synonyms and adjectives produces diminishing returns. If you’ve been prompting since V4, you have habits to unlearn. Do it now.

2. Turn on Raw mode when you want photography

V7’s default aesthetic is gorgeous, but it’s processed. There’s a built-in artistic hand on the dial. When you want something that looks like an actual photograph instead of a beautifully rendered image, kill it.

Turning on Raw mode disables Midjourney’s built-in aesthetic processing, which can make outputs look more naturally photographic rather than stylised. As Daniel Riley teaches: “If you use Raw, it can actually be good for giving you more realistic results because it turns off the artistic component of Midjourney.” Pair Raw mode with a low stylize value and you’ll see the difference instantly. The image stops trying to be art and starts trying to be a photo.

3. Learn what —s actually does (and stop leaving it on default)

The stylize parameter is the most misunderstood knob in Midjourney. People crank it because “more style = more good,” and then wonder why their product shot suddenly looks like a fantasy painting.

Here’s the real map: —s 0–150 means V7 tries to render your description as literally as possible. Accurate, sometimes flat. —s 700–1000 means the model interprets your prompt creatively, often beautifully, often at the expense of your specific requests. Most professional work lives in —s 200–400. Push higher for illustration and abstract work, stay lower for product and documentary.

If you’re shooting a real-looking product, a real-looking person, or anything documentary, you want low stylize. If you’re making concept art, push it up. That’s the whole rule.

4. Use —sref and —oref instead of describing a style

This is the V7 superpower people aren’t using enough. Trying to describe “the exact muted teal-and-orange palette of a Wes Anderson film” with words is a losing game. Pointing the model at a reference image is not.

—sref and —cref replace style-keyword spam. In V6, achieving a consistent look meant cramming your prompt with style adjectives: “film noir, high contrast, moody chiaroscuro, 1940s detective aesthetic.” In V7, you can attach a reference image URL with —sref [URL] and the model reads that visual directly.

For character consistency, same face across ten images, V7 introduced something even better. In Midjourney V7, achieve consistent image generation with —oref (Omni-Reference) to input a reference image via URL or web UI upload. After the image is uploaded, you can control how the reference image influence the final creation using —ow (Omni-Weight), a value from 0 to 1000 (default is 100). Lower —ow values allow for more stylistic interpretation, ideal for transferring styles. With higher value, the created image will have closer replication of details like faces or clothing.

Translation: if you need the same character in five different scenes, —oref with a high —ow is your friend. If you want a vibe transferred but not a literal copy, drop the weight. This is the workflow you should be building around.

5. Name your failure modes with —no

Look at your last batch of generations. There’s a thing that keeps showing up that you didn’t ask for, isn’t there? A weird hand in the corner. A logo on a shirt. A lens flare you don’t want. Stop hoping it’ll go away. Name it and ban it.

Use —no to name your recurring failure modes. Look at your last ten results. What keeps appearing that you don’t want? Watermark artifacts? Background text? An unnecessary lens flare? A prop you never asked for? Name it in —no. In V7, this parameter is more reliable than earlier versions. —no text, watermark, crowded background saves regenerations.

This one habit will save you more credits than any other on this list. Keep a running —no list for the kind of work you do most and paste it onto every prompt.

6. Iterate one variable at a time, not five

This is the discipline thing. You generate four images, you don’t love them, and the temptation is to rewrite the entire prompt. Don’t. You’ll never learn what actually changed the result.

Create a “base prompt” you’re happy with, then duplicate it and change only one parameter at a time. This lets you see exactly what each adjustment does to your output. Don’t change everything at once. Test one parameter at a time. Bump —s up by 100, or swap —ar, or add a single —sref. See what moved. Then move the next thing.

And don’t try to write the perfect prompt on the first attempt. “The workflow shift that saved me the most time: stop trying to write the perfect prompt on the first try. Short prompt, pick the best of 4, then add detail to that direction. You get better results in half the generation credits.” That’s the whole job. Short prompt, pick the best direction, then push.

A bonus, because it matters: use /describe

Stuck staring at an image you love and wondering how to describe its style to V7? Don’t guess. Upload any image into Midjourney’s /describe command and the model will generate four prompt interpretations of what it sees.

You’ll get four prompts back, written in V7’s own vocabulary. Steal the phrasing. It’s the fastest way to learn how the model thinks about light, mood, and composition, and it’s free.

The one habit that ties it all together: treat Midjourney like a camera you’re learning, not a slot machine you’re pulling. Every parameter does something specific. Every reference image teaches the model something concrete. The people getting great images aren’t lucky and they aren’t gifted. They’re just paying attention to what each knob does. Start doing that, and your hit rate goes up overnight.