Midjourney Negative Prompts: What to Exclude for Better AI Images

I spent three hours once trying to get a clean product shot out of Midjourney. No watermarks. No blurry edges. No mysteriously extra finger appearing on the hand holding the product. Every iteration had at least one of these problems, and I kept trying to fix them by rewriting the positive prompt.

That was the mistake. I was trying to describe my way to quality when I should have been excluding my way to it.

Negative prompts — specifically Midjourney's --no parameter — don't get enough attention. Most guides spend 95% of their time on how to describe what you want. The exclusion side of the equation is where a lot of the quality control actually lives.

This guide covers what to exclude, why certain things need excluding, and the specific --no strings that actually work.

How Midjourney's --no Parameter Works

The --no parameter tells Midjourney's model to steer the generation away from specific concepts. It works by applying negative conditioning to those terms during the diffusion process — mathematically, it's the opposite of prompting for something.

The syntax is simple:

/imagine your main prompt here --no thing1, thing2, thing3

What you're doing is telling the model: generate everything in my positive prompt, but actively move away from these excluded concepts during the image generation process.

A few things to understand about how it actually works:

It's probabilistic, not absolute. Strong stylistic associations don't always disappear completely. If your positive prompt has heavy painterly vibes and you add --no painting, you'll reduce but not necessarily eliminate that quality. The more specific your exclusion, the better.

Order doesn't matter within --no. Whether you write --no watermark, text or --no text, watermark makes no meaningful difference.

You can combine it with other parameters. --no text watermark --style raw --ar 16:9 all work together.

The Universal Baseline --no String

Before you customize anything, start here. This combination eliminates the most common artifacts across virtually every generation type:

--no text, watermark, logo, signature, border, frame, blurry, 
blur, out of focus, jpeg artifacts, compression artifacts, 
low quality, bad quality, worst quality, duplicate, cropped

Add this to every prompt as a starting point. Then layer on specific exclusions for your use case.

Anatomy and Human Figure Problems

This is the most famous category of Midjourney failures. Human hands, faces with multiple eyes, bodies with extra limbs — these come from how diffusion models learn from imperfect training data where human anatomy isn't always clearly visible.

For any portrait or figure shot, use this anatomy block:

--no extra fingers, extra limbs, deformed hands, bad hands, 
bad anatomy, mutated, deformed, ugly, disfigured, 
fused fingers, missing fingers, extra faces

The --quality 2 parameter also helps here — more generation steps generally produce more anatomically coherent results, though at double the credit cost.

Controlling Style Bleed

One of the most useful and underused applications of negative prompting. Style bleed happens when unwanted aesthetic influences contaminate an otherwise clear style direction.

When you want photorealistic but keep getting illustrated:

--no illustration, painting, artwork, drawn, sketch, anime, 
cartoon, render, CGI, digital art, concept art, painting style

When you want clean illustration but keep getting photorealistic:

--no photograph, photo, photography, realistic, photorealistic, 
real, camera, lens, DSLR

When you want minimalist/flat design but keep getting textured:

--no texture, grungy, noise, grain, rough, vintage, aged, 
weathered, distressed

When you want editorial/clean corporate but keep getting fantasy:

--no fantasy, magical, ethereal, mystical, surreal, dreamlike, 
glowing, neon, sci-fi

The Midjourney V6 parameters guide covers the full range of style parameters you can combine with these negative strings.

Lighting and Atmosphere Control

Bad lighting is the second-biggest quality killer after anatomy issues. Midjourney has strong tendencies toward dramatic, high-contrast, and sometimes harsh lighting that doesn't suit every image type.

For clean, professional product photography:

--no harsh shadows, dark shadows, overexposed, blown out highlights, 
underexposed, dark, moody, dramatic lighting, horror lighting

For soft portrait lighting (eliminating the harsh flash look):

--no flash photography, harsh light, hard shadows, raccoon eyes, 
unflattering lighting, overhead lighting

For bright, airy lifestyle imagery:

--no dark background, black background, dark mood, night, 
dim lighting, shadow-heavy, noir

For flat, even product backgrounds:

--no gradient background, textured background, busy background, 
patterned background, distracting background

Eliminating Unwanted Text and Graphics

Midjourney has an annoying habit of adding fake text — placeholder labels, imaginary watermarks, random lettering — especially in scenes that look like they should have text (storefronts, product packaging, signs).

--no text, words, letters, numbers, typography, font, 
written text, calligraphy, handwriting, label, caption, 
subtitle, watermark, logo, signature, copyright symbol

This is one of the most reliable negative prompt applications. The exclusion is nearly 100% effective for generated watermarks and random text overlays. It's less effective if your scene strongly implies text (a bookshelf, a street sign in a realistic city shot) — in those cases, you also need to adjust the positive prompt to avoid the context that triggers text generation.

Background and Composition Control

For clean product shots with white or solid backgrounds:

--no complex background, busy background, distracting background, 
outdoor background, people in background, cluttered background, 
furniture, objects, detailed background

For portraits where the subject should fill the frame:

--no wide shot, full body, zoomed out, distant, small subject, 
large background proportion

For avoiding the typical "AI art composition":

--no centered composition, symmetrical, perfect symmetry, 
AI composition, generic pose, stock photo pose

That last one deserves a note. There's a recognizable "AI aesthetic" in composition — perfectly centered subjects, very symmetrical framing, specific ways of arranging elements that appear over and over. If you're trying to make your images look less AI-generated, breaking compositional patterns with negative prompts (combined with stronger positive composition direction) genuinely helps.

Category-Specific --no Templates

These are ready to copy. Adjust the positive part of your prompt freely — these negative strings are designed to work as defaults for each category.

Portrait photography:

professional headshot of [person description], natural lighting, 
shallow depth of field --no extra fingers, bad hands, bad anatomy, 
deformed, ugly, blurry, watermark, text, harsh shadows, extra faces,
multiple people, clone, duplicate --ar 3:4 --q 2

Product photography:

[product] on white background, studio lighting, 
sharp detail, professional product photo --no text, watermark, 
logo, shadow artifacts, reflection artifacts, dust, scratches, 
busy background, props, hands --ar 1:1 --style raw

Architectural / interior design:

[interior space description], professional architecture photography, 
natural light --no people, figures, humans, furniture clutter, 
blurry, distortion, fisheye, wide angle distortion, 
perspective distortion --ar 16:9

Fantasy illustration:

[scene description], fantasy digital art, detailed illustration --no 
photograph, realistic, photorealistic, text, watermark, signature, 
ugly, low quality, amateur, out of focus --ar 16:9

Logo / icon design:

[logo concept], clean vector logo, minimal, flat design, 
professional --no 3D, shadow, gradient, complex background, 
photograph, realistic, hands, people, cluttered, text (unless 
text is part of your intended logo)

For beginners still building their prompting intuition, the Midjourney prompts for beginners guide covers the positive prompting foundations that these negative strings work alongside. And the notes on prompt engineering has quick reference syntax for both positive and negative approaches.

When --no Doesn't Work (And What to Do Instead)

Negative prompts are not magic. There are failure cases worth understanding.

Problem: The excluded element keeps appearing despite --no

This happens when the excluded concept is strongly embedded in the subject matter. If you're generating a library scene and adding --no books, you're fighting the model's strong association between the concepts. The fix is to change your positive prompt to describe the scene differently, reducing the context that triggers the unwanted element.

Problem: --no is pulling the image toward the opposite extreme

Telling Midjourney --no dark, moody sometimes overcompensates into blown-out, overexposed brightness. The negative parameter moves away from one extreme but doesn't know where to stop. Add positive lighting direction: soft natural lighting, diffused light --no dark, moody, harsh shadows.

Problem: Style exclusions are making the image look generic

If you're excluding too many styles at once (--no painting, illustration, anime, cartoon, sketch, artwork), you may be removing so many aesthetic options that the model defaults to the blandest photorealistic interpretation. Be selective. Exclude what's actually appearing, not every style you don't want.

Testing and Iterating --no Strings

The most efficient approach is iterative. Don't build massive --no strings from scratch.

1. Generate with your positive prompt alone
2. Identify the specific problems in the output
3. Add only the --no terms that address those specific problems
4. Regenerate and assess

Most reliable quality improvements come from targeted exclusions of actual problems, not preemptive lists of every bad thing that could theoretically appear.

The prompt engineering basics quiz is worth taking if you want to check whether you understand the mechanics of how positive and negative prompting interact — a lot of the quiz questions are directly applicable to Midjourney.

One pattern I've found consistently useful: keep a personal --no library organized by image type. After a few weeks of generating in a particular style, you'll know exactly which exclusions your use cases need. Building those defaults saves significant iteration time on every new generation.