Why Your Midjourney Prompts Don't Work (10 Common Mistakes, 2026)

TL;DR: Most "bad" Midjourney output traces to one of 10 specific mistakes, and each has a roughly 30-second fix. The catch is that broken prompts usually fail on two or three mistakes at once, so the output looks mysteriously wrong rather than wrong in one obvious way. This guide gives you the diagnostic, the fixes, and copy-pasteable before/after rewrites for Midjourney v7.

Why are my Midjourney prompts not working?

Your Midjourney prompts are not working because of a small set of repeatable mistakes — usually a missing --raw flag on photo prompts, a vague subject, too many stacked style words, or a --stylize value wrong for the task. Midjourney v7 is more prompt-faithful than older versions, so the fix is almost always editing the prompt, not the tool. Most broken prompts trip over two or three of these at once.

That last sentence is the part people miss. When a single image looks "off," it is tempting to blame the model, regenerate, and hope. But Midjourney v7 — which became the default model in June 2025 — is genuinely good at following clear instructions. It renders bodies, hands, and objects with far more coherence than v6, and it handles text and image prompts with stunning precision when the prompt is well-formed. So when output disappoints, the prompt is the variable to interrogate first.

This article walks through the 10 mistakes that account for the overwhelming majority of disappointing Midjourney results, gives each a concrete fix, and ends with a 30-second diagnostic checklist you can run on any failing prompt. Everything here is written for v7 (and forward-compatible with the v8.1 model line), so the parameters and defaults match what you are actually typing today, not what worked in 2024.

Mistake 1: Missing --raw on photographic prompts

Symptom: photo prompts come out slightly painterly, oversaturated, or "AI-glossy" instead of looking like real photography.

Why it happens: Midjourney applies a built-in house aesthetic by default — a cinematic, slightly stylized flair that is gorgeous for concept art and wrong for documentary realism. The --raw parameter (technically --style raw) tells Midjourney to suppress that built-in artistic bias and interpret your prompt more literally, producing cleaner, flatter, more photographic results.

The fix: add --raw and lower --s to the 100-150 range for photo realism.

Before: A 30yo woman, wool coat, golden hour --s 500 --v 7
After:  A 30yo woman, wool coat, golden hour --s 150 --raw --v 7

This single change is responsible for more "wow, that fixed it" moments than any other on this list. If you only remember one thing from this article, remember that photo prompts without --raw are fighting the model's default personality. The house aesthetic grew stronger in v7, which is exactly why --raw matters more now than it did two versions ago.

A useful mental model: think of --raw as turning off Midjourney's opinion and --stylize as turning up its opinion. For a believable photograph you want the opinion low and off. For a poster or album cover you want it high and on.

Mistake 2: Using a generic subject

Symptom: faceless, forgettable, stock-photo-looking output. The image is technically fine but has no character.

Why it happens: a phrase like "a woman" matches an enormous cluster of training data, so the model averages across millions of examples. Averages are bland by definition. The more generic your subject, the more generic the result.

The fix: give the subject three to five specific descriptors — hair, clothing, a distinguishing feature, and an expression or action.

Before: a woman walking through Paris
After:  a 30-year-old woman with curly red hair and light freckles,
        wearing a charcoal wool coat, walking briskly through a
        rain-slicked Paris cobblestone street, looking back over
        her shoulder

Specificity is the cheapest quality upgrade in image generation. You are not making the prompt longer for the sake of length — every descriptor narrows the model toward a single, coherent person instead of a statistical blur of everyone. If you want to systematize this, keep a small library of reusable subject blocks with {{variables}} you can swap in, which is exactly what the save-and-reuse prompt library inside Prompt Architects is built for.

Mistake 3: Stacking five or more style modifiers

Symptom: muddy, indistinct output that doesn't commit to any single look.

Why it happens: when you stack cinematic, dramatic, atmospheric, moody, ethereal, golden hour, anamorphic, you give the model seven partly-conflicting directions. It cannot satisfy all of them, so it averages — and the average of seven styles is mush.

The fix: two to three modifiers maximum, each from a different category (medium + lighting + era, for example). Pick the strongest word in each category instead of piling on synonyms.

Before: cinematic, dramatic, atmospheric, moody, ethereal, golden hour, anamorphic
After:  35mm film, golden hour, anamorphic lens flare

Notice that "cinematic," "dramatic," "atmospheric," and "moody" are nearly synonyms — they all point at the same vibe, so listing all four adds noise without adding signal. The fix keeps one medium (35mm film), one lighting cue (golden hour), and one optical detail (anamorphic lens flare). Three orthogonal instructions the model can actually execute.

Here is a quick category map to help you choose non-overlapping modifiers:

Choose one from a few categories, not five from one. Two or three well-chosen modifiers consistently beat seven competing ones.

Mistake 4: Forgetting --ar (aspect ratio)

Symptom: you get a square image when you wanted a vertical story or a widescreen banner.

Why it happens: Midjourney outputs 1:1 by default. If you don't specify an aspect ratio, you get a square — every time.

The fix: always include --ar. Match it to where the image will actually live:

Instagram Stories / TikTok / Reels (vertical):  --ar 9:16
Instagram or LinkedIn feed (portrait):          --ar 4:5
YouTube thumbnail / blog header (landscape):    --ar 16:9
Cinematic widescreen:                           --ar 21:9
Pinterest:                                        --ar 2:3
Print poster (portrait):                         --ar 2:3 or --ar 3:4

This is a low-skill, high-impact habit. Cropping a square down to 9:16 in post throws away half your composition and usually decapitates your subject. Setting --ar up front lets the model compose for the frame — leaving headroom, balancing negative space, and placing the subject where it belongs. If you generate for multiple platforms regularly, decide the target ratio before you write the prompt, not after.

Mistake 5: Mixing framing instructions

Symptom: a confused composition that fits no single shot type — the camera seems to be in two places at once.

Why it happens: phrases like "wide shot close-up" or "medium shot extreme close-up" send the model contradictory framing signals. A wide shot and a close-up are mutually exclusive; the model splits the difference into something incoherent.

The fix: pick exactly one framing instruction per prompt.

Before: wide shot close-up portrait
After:  medium close-up portrait

Cinematographers use a precise vocabulary of shot sizes, and Midjourney understands it well when you use one term cleanly. Here is the ladder, from widest to tightest:

• Extreme wide shot — subject tiny in a vast environment
• Wide / establishing shot — full subject plus surroundings
• Full shot — whole body, head to toe
• Medium shot — waist up
• Medium close-up — chest up (the workhorse for portraits)
• Close-up — head and shoulders
• Extreme close-up — eyes, lips, or a single detail

Choose one rung. If you want both a wide context and a tight detail, that is two separate generations, not one prompt.

Mistake 6: No lighting cue

Symptom: flat, generically lit output with no depth or mood.

Why it happens: lighting is arguably half of what makes an image feel intentional. With no lighting instruction, the model defaults to something safe and even — which reads as "nobody chose this."

The fix: specify the source, the direction, and the mood.

Before: portrait of a man
After:  portrait of a man, golden hour warm light from the west,
        soft side rim light, faint atmospheric haze

Lighting modifiers that behave reliably in v7:

• Time / source: golden hour, blue hour, harsh midday sun, candlelight, neon, moonlight
• Setup: studio softbox, ring light, rim light, backlit, side-lit, top-down
• Mood: dramatic chiaroscuro, soft diffused, high-key, low-key, mixed warm and cool

Combine one source with one direction and, optionally, one mood word: "candlelight from below, low-key, dramatic chiaroscuro." That is three coordinated lighting instructions, not three competing styles — they reinforce each other instead of canceling out.

Mistake 7: Using the wrong --stylize value for the task

Symptom: a photo that looks too painterly, or an illustration that looks too flat and literal.

Why it happens: --stylize (or --s) controls how much creative freedom Midjourney has to transform your prompt. The default is 100 and the range runs 0 to 1000 — at low values the model follows your words closely and stays realistic; at high values it experiments boldly, adding flair and drama at the cost of literal accuracy.

The fix: tune --s to the job. This table reflects values that hold up consistently in v7:

The mistake is treating --s as a "quality" dial — it is not. A high --s does not make a more correct image, it makes a more opinionated one. If you crank stylize on a product shot expecting "better," you get a product that looks subtly invented. Match the dial to the intent: low for faithful, high for expressive. For a deeper walk through every parameter and how they interact, see our Midjourney parameters cheat sheet.

Mistake 8: Cramming too many subjects into one prompt

Symptom: subjects merge into each other, key details vanish, or one figure dominates and the rest become blobs.

Why it happens: Midjourney handles one or two main subjects well. Beyond that, its attention spreads thin and coherence collapses — five clearly-described people in one frame become a crowd of half-rendered approximations.

The fix: focus on a single primary subject per prompt and build complex compositions from multiple generations.

Before: Five people of different ages and backgrounds gathered around a
        table, each holding a different object, in a busy market

After:  A 60-year-old market vendor in a canvas apron arranging fresh
        produce on a wooden crate, soft morning light, busy market
        softly blurred in the background
        (generate other figures separately, then composite if needed)

The "after" prompt gives one subject the full weight of the model's attention, which is why the vendor reads as a real, specific person. If your scene genuinely needs five people interacting, that is a compositing job — generate the hero element cleanly, then bring in supporting elements through separate generations or an editing pass. Asking one prompt to nail five distinct, posed, prop-holding figures is asking for the muddle you get.

Mistake 9: Asking for impossible specificity

Symptom: the model flatly ignores precise details you carefully specified — exact counts, exact text, exact object placement.

Why it happens: diffusion image models have real, structural limits. They cannot reliably count above roughly four or five objects, cannot render specific text longer than about three words without misspelling, and cannot place objects with surgical precision. These are not prompt-writing failures; they are properties of how the models work.

The fix: simplify, or work in stages.

Avoid:  "Exactly seven red apples in a precise pyramid arrangement"
Better: "A pile of red apples in a wooden crate"

Avoid:  "A storefront sign reading 'Madeleine's Artisanal Bakery & Cafe'"
Better: "A storefront sign reading 'Bakery'" (short, in quotes)

A few practical rules of thumb for working around model limits:

• Counting: don't specify exact numbers above four. Generate "a pile" or "several" and select the variant that happens to land where you want.
• Text: keep it to one to three words, wrap it in quotes, and use a common font. For anything text-heavy — packaging, posters with paragraphs, UI mockups — switch to Ideogram, which is purpose-built for typography.
• Precise placement: describe relationships loosely ("in the foreground," "to the left") rather than exact coordinates, or use the editor and regional prompting to place things deliberately.

Knowing what not to ask for is as valuable as knowing what to ask for. Half of "Midjourney is broken" complaints are really "I asked for something diffusion models can't do."

Mistake 10: Never adjusting --chaos for variation

Symptom: your four generations look nearly identical, so there is nothing meaningful to choose between. Or the opposite — you wanted a consistent series and got four wildly different images.

Why it happens: --chaos (or --c) controls how different the four results in a grid are from one another. The default is 0 and the range is 0 to 100. At low chaos you get four variations on one idea; at high chaos you get four genuinely different interpretations (with looser adherence to the prompt).

The fix: set --c deliberately based on whether you are exploring or refining.

Exploring a new concept:        --c 25-30
Building a consistent series:    --c 0-10
Wild experimentation:            --c 50+ (mostly throwaway, occasional gold)

Early in a project, when you don't yet know what you want, crank chaos to 25-30 so each generation shows you a different direction. Once you have found the look, drop it back toward 0-10 to produce a coherent set. Leaving chaos at its default of 0 for everything means you keep regenerating slight variations of the same idea and wondering why nothing new appears.

The 30-second Midjourney diagnostic checklist

When output is bad, don't regenerate blindly. Run the prompt through these 10 checks. Each maps directly to one mistake above.

Score it. If your prompt passes fewer than 7 of these 10 checks, the prompt — not Midjourney — explains the bad output. Hit 7 or more before you generate and the quality lift is consistent and obvious. This is the single highest-leverage habit in this whole article: spend 30 seconds scoring before you spend a generation.

A full before/after rewrite

Here is a deliberately broken prompt that commits most of the mistakes at once:

woman walking in city, cinematic, dramatic, atmospheric, moody, beautiful,
detailed, masterpiece, 4k, photorealistic, ultra-realistic, perfect

What's wrong with it:

• 11 modifiers, most of them redundant synonyms ("cinematic/dramatic/atmospheric/moody," "photorealistic/ultra-realistic")
• No specific subject — just "woman"
• No lighting cue
• No aspect ratio (so you'll get a square)
• No --raw on what is clearly meant to be a photo
• No parameters at all — stylize, chaos, version all left to default
• Junk tokens like "masterpiece," "4k," and "perfect" that add nothing in v7

Now the same idea, with every fix applied:

A 30-year-old woman with curly red hair, wearing a charcoal wool coat,
walking briskly across a wet cobblestone street in Paris at autumn dusk,
light rain falling.

Medium tracking shot, 35mm lens, slight low angle.
Golden hour warm light from the west mixing with cool blue from streetlamps.
Anamorphic lens flare.

--ar 21:9 --s 200 --raw --v 7

Same core concept. The difference in output is night and day — the second prompt produces a specific, well-lit, properly-framed cinematic still, while the first produces a generic AI woman in a square frame. Nothing here is exotic; it is just the 10 fixes applied in sequence.

How do I keep a character consistent across Midjourney images?

This is the single most-asked Midjourney question, and the answer changed with v7 — which trips up anyone working from older guides.

In v6, character consistency ran on --cref. In v7, that system was replaced by Omni Reference (--oref), which works not just for characters but for objects, vehicles, and creatures from a single reference image. The strength of the reference is controlled by the omni-weight parameter --ow, which ranges from 0 to 1000 and defaults to 100.

Subtle style change (photo to anime):     --oref [URL] --ow 25
Balanced consistency (default-ish):        --oref [URL] --ow 100
Strong face/outfit lock:                   --oref [URL] --ow 400

A few things to know that the docs make explicit:

• Lower --ow (25-50) when you want to change the style — for example, turning a photo of a person into an anime version. High weight would fight the style change.
• Higher --ow (around 400) when you need a face or specific clothing preserved exactly across scenes.
• --stylize and --exp compete with the reference. If you run high stylize, you need a correspondingly higher --ow to keep the reference fidelity, because they pull against each other.

Used well, Omni Reference delivers strong character consistency across poses, outfits, and environments from one image — no elaborate workflow required. If your "character keeps changing" prompts are copied from a 2024 tutorial, the silent failure is almost always that you are still typing --cref into a model that no longer treats it as the primary character system.

What changed in Midjourney v7 (and why old prompts break)

If a prompt that worked in 2024 suddenly produces worse results, the model underneath it changed. Midjourney retrains with each major version, so modifier weights and defaults shift. Here is what actually moved with v7, which launched in alpha on April 3-4, 2025 and became the default on June 17, 2025:

Two of these deserve emphasis. First, personalization is on by default in v7 — it is the first Midjourney model to ship with personalization enabled, which you unlock after rating a series of images. This is great for getting "your" look, but it also means two people running the identical prompt can get different results, so don't assume a prompt is broken because it doesn't match someone else's screenshot.

Second, Draft Mode runs at half the cost and roughly 10x the rendering speed, trading some quality for velocity. It is the right tool for the exploration phase of the diagnostic — burn through directions fast in draft, then enhance the winner to full quality. For a structured walkthrough of building a repeatable workflow around these features, see our Midjourney workflow guide.

When should you stop iterating and switch approach?

Iterating a fundamentally broken prompt six times costs more time and credits than starting over. Use a two-strike rule: if the same prompt fails twice after you have applied the diagnostic checklist, stop iterating and change one of three things.

1. Switch tools. Text-heavy image? Move to Ideogram. Anime or manga style? Use Niji, which is tuned separately from the main model. Each tool has a lane; fighting Midjourney to do Ideogram's job wastes both.
2. Switch approach. If text-to-image keeps missing, move to image-to-image with a reference, or use Omni Reference to pin the element you can't get to hold.
3. Break the task. Multi-subject scene? Generate each element cleanly on its own and composite. One prompt cannot reliably do what three prompts plus an edit can.

The skill here is recognizing category failures versus tuning failures. If two diagnostic-clean attempts both fail, you are probably asking for something in the "impossible specificity" bucket — and no amount of rerolling fixes a structural limit.

Power moves for reliable Midjourney output

Once the basics are solid, these habits compound:

1. Template your winners. Save your 10 best prompts as templates with {{subject}}, {{lighting}}, and {{ar}} placeholders so you can produce consistent results without retyping. A reusable prompt library turns a lucky generation into a repeatable asset.
2. A/B test one parameter at a time. Run the same prompt at --s 150 and --s 250, change nothing else, and note which works for which task. Changing two things at once tells you nothing.
3. Pair --oref with a fixed look for character series. Lock the face with a higher --ow, then vary scene and pose around it.
4. Default to --c 20-30 while exploring, then drop to --c 0-10 once you've found the direction. Make chaos a conscious choice, not a forgotten default.
5. Keep a "tested combinations" list. Pairings that reliably work for your style — your go-to medium plus lighting plus stylize value — become your personal preset bank.

For broader prompt-craft principles that apply beyond Midjourney, our guide on writing prompts that actually work across AI tools covers the structure-first mindset that underlies all of this.

What to do next

You don't need to memorize all 10 mistakes today. Do this instead:

1. Pick your worst recent generation. Run it through the 10-point diagnostic checklist above and score it honestly.
2. Rewrite incorporating every failed check, then regenerate.
3. Note which single fix produced the biggest lift. That fix is your habitual weakness — address it first in every future prompt and your average quality climbs on its own.

The parameters and defaults will keep shifting as Midjourney ships new versions, but the diagnostic transfers. Specific subject, controlled modifiers, intentional lighting, the right --stylize, one framing, sane scope, and parameters chosen on purpose — that is the whole game. Tools that surface Midjourney's parameters in a UI with cinematic presets (like Prompt Architects) remove the typing friction, but the thinking above is what actually moves the output.

Frequently asked questions

Why do my Midjourney images look generic? Three usual culprits: missing --raw on photo prompts (Midjourney applies its house aesthetic by default), a generic subject like "a woman" instead of a specific one, and stacking five or more style modifiers that average to mush. Fix any one and quality jumps.

Why doesn't Midjourney render text correctly? V7 improved text rendering but still misspells phrases longer than about three words and distorts uncommon fonts. For text-critical images, switch to Ideogram. For Midjourney text, keep phrases short, wrap them in quotes, and use familiar fonts.

Why are my Midjourney characters inconsistent across images? In v7, character consistency runs on Omni Reference (--oref) with the omni-weight --ow (range 0-1000, default 100). Paste your reference image URL after --oref and raise --ow toward 400 to lock a face and outfit. Without a reference, the model reinvents the character every time.

Why do my Midjourney prompts produce muddy or muted colors? Usually --stylize is too low or there's no palette anchor. The default --s 100 is neutral; --s 250-500 produces saturated, branded looks. Add explicit palette modifiers like "jewel tones" or "warm gold and cool blue contrast."

What is the best stylize value in Midjourney v7? It depends on the task: --s 50-150 with --raw for photorealism, 150-250 for editorial photography, 400-700 for stylized illustration, 750-1000 for strong artistic interpretation. The default is 100 and the range is 0 to 1000.

Why does the same prompt produce different styles across Midjourney versions? Each major version retrains the model, shifting modifier weights. V6 was more painterly; v7 is more photoreal and prompt-faithful, and turns on personalization by default. Re-test saved prompts after every major version.

Does Midjourney v7 still use --cref for character reference? No. V7 replaced the v6 --cref system with Omni Reference (--oref), which works for characters, objects, vehicles, and creatures. Prompts copied from 2024 guides using --cref may silently fail.

How many style modifiers should a Midjourney prompt have? Two to three, from different categories such as medium, lighting, and era. Stacking five or more forces the model to average between them, producing muddy output.

By Nafiul Hasan — Founder of Prompt Architects and a daily Midjourney user who has tested thousands of prompts across v6, v7, and the v8 model line. Last updated: June 10, 2026.