Negative Prompts: What to Tell AI Not to Generate
Learn when negative prompts help AI image generation, how to write specific exclusions, and when a positive rewrite works better than a long negative list.
Negative Prompts: What to Tell AI Not to Generate
There’s a strange moment that happens when you start generating images with AI.
You describe what you want. You add the subject, the lighting, the camera angle, the mood. You press generate. And somehow the image still comes back with something you never asked for.
Extra fingers. Plastic-looking skin. A random object in the corner. A background that doesn’t belong there. Lighting that’s suddenly doing far too much. A face that looks almost right, but not quite.
So you add more words. Then more. Then a long list of things the model should avoid: no text, no watermark, no extra fingers, no distortion, no blur, no bad anatomy, no weird hands, no oversaturation, no cartoon effect.
The prompt gets longer. The result doesn’t necessarily get better.
This is where negative prompting becomes useful — and also where it’s most often misunderstood. A negative prompt isn’t a list of everything you dislike. It’s a way of constraining the visual space around your idea, and that difference matters more than it sounds like it should.
What is a negative prompt?
A negative prompt is an instruction describing what shouldn’t appear in the result. Take this main prompt:
A cinematic portrait of a woman standing in an abandoned train station at night, soft overhead lighting, realistic skin texture, 35mm photography.
A negative prompt for it might be:
text, watermark, logo, distorted hands, extra fingers, plastic skin, oversharpening, excessive saturation
The main prompt defines the destination. The negative prompt closes off some of the roads you don’t want the model taking to get there. A positive prompt says go here. A negative prompt says not that way — and sometimes that second instruction is exactly what stops the machine wandering off into the weeds somewhere between your intention and its accident.
Not every tool handles this the same way, worth knowing before you go further: Midjourney and Stable Diffusion both have dedicated negative-prompt support (Midjourney’s –no parameter, Stable Diffusion’s separate negative field), while OpenAI’s GPT Image models don’t have an equivalent parameter at all — exclusions there have to be folded into the ordinary, positive-language prompt instead (“a clean empty street” rather than “no cars”). Check which kind of tool you’re using before you copy a workflow from one to the other.
Negative prompting is not the same as “make it better”
The most common mistake is writing vague negative prompts:
bad quality, ugly, weird, terrible, low quality, bad image
What does the model actually learn from that? Very little. “Bad” isn’t a visual property. “Ugly” isn’t a precise instruction. Instead of “bad anatomy,” try “distorted hands, incorrect finger count, malformed limbs.” Instead of “bad lighting,” try “blown highlights, crushed shadows, harsh frontal flash.” Instead of “bad skin,” try “plastic skin, excessive smoothing, artificial skin texture.” The more specific the unwanted behaviour, the more useful the instruction.
A useful habit: stop thinking in adjectives and start thinking in failure modes. Ask what could actually go wrong with this particular image. For a portrait, that’s things like distorted hands, unnatural eyes, duplicated facial features or incorrect proportions. For a product shot, it’s floating objects, warped packaging, unreadable labels or inconsistent shadows. For architecture, it’s impossible structures, warped perspective or duplicated elements. That’s a far more useful starting point than “ugly, bad, low quality, weird” — you’re not telling the model the image is bad, you’re telling it which specific mistakes to avoid.
Objects, artifacts, composition, lighting
A few of the most common uses:
Unwanted objects. If a minimalist gallery scene keeps sprouting people, furniture or cars you never asked for, “people, furniture, vegetation, vehicles, signage, text” is a concrete, useful negative prompt. Just don’t pre-emptively list forty objects that aren’t actually appearing — negative prompting should respond to real problems, not hypothetical ones.
Recurring visual artifacts. Distorted anatomy, duplicated objects, malformed hands, warped text and unnatural reflections are the classic failure patterns across most image generators. “Distorted hands, extra fingers, duplicated limbs, malformed facial features” tells the model what to avoid. “Bad anatomy, ugly hands” tells it your frustration, which it doesn’t have much use for.
Composition. Negative prompts can protect a layout, not just remove objects. If you need a subject in the left third of the frame with room for text on the right, state that positively first — “subject positioned in the left third of the frame, generous negative space on the right” — then use the negative prompt to guard against the model’s favourite default: “centred composition, subject filling the entire frame, cropped subject.” The positive prompt builds the design. The negative prompt protects it from drifting back to what the model would otherwise default to.
Lighting. If you want soft directional light and restrained highlights but keep getting glowing edges and blown-out backgrounds, “excessive bloom, blown highlights, harsh contrast, artificial lens flare, overexposure” can help — but be careful not to strip out so many lighting qualities that the image goes flat. A negative prompt should protect the lighting concept, not suffocate it.
Be careful with style negatives
Say you want a realistic photograph and you’re tempted to write “cartoon, anime, illustration, painting, 3D render, CGI, digital art” as a blanket exclusion. That can help if the model keeps drifting toward one of those styles specifically. But the model isn’t reading your negative prompt like a human art director reading a brief — it’s working with relationships between visual concepts, and throwing a dozen competing style terms at it can make an image less coherent rather than more realistic. Midjourney’s own documentation warns about a related, sharper version of this problem: its –no parameter reads each word independently, so “–no modern clothing” can get parsed as “no modern” plus “no clothing” — occasionally tripping a content warning for something you never intended. If your tool works that way, describe the clothing you do want rather than excluding it.
If the result keeps looking like a 3D render specifically, target that symptom directly — “plastic surfaces, artificial reflections, CGI appearance, overly perfect geometry” — rather than banning every style you can think of.
Protecting realism
Realism isn’t just the absence of illustration — it often comes down to small imperfections. Instead of “make it realistic” plus “no unrealistic things,” think about what actually signals artificiality. For a portrait: “plastic skin, excessive facial symmetry, excessive skin smoothing, artificial eyes, over-retouched appearance.” For an environment: “perfectly clean surfaces, identical repeated objects, unnatural symmetry, artificial reflections.” For a photograph: “excessive HDR, oversharpening, excessive clarity, unnatural bokeh, artificial lens effects.” These describe the actual signals of artificiality instead of just asking for “realism” and hoping.
Don’t repeat the same idea ten different ways
A negative prompt like “blurry, blur, blurred, low resolution, poor quality, bad quality, low detail, lack of detail, fuzzy, unclear” feels thorough. It usually isn’t — most of those phrases describe the same underlying problem, so you’re spending prompt space without adding information. “Motion blur, excessive softness, low detail” covers the same ground in three distinct, useful constraints. The goal isn’t an impressive-looking list. It’s a precise one.
Don’t use negative prompts to fix a bad positive prompt
This might be the most important rule here. If your positive prompt is “a person in a city” and your negative prompt is “no crowds, no buildings, no vehicles, no signs, no advertisements, no streets, no urban clutter,” you’ve built a contradiction — the model has almost nowhere left to go. When the desired scene isn’t clearly defined, negative prompting turns into a desperate attempt to repair it after the fact. Fix the positive prompt first: describe the subject, environment, composition, lighting, camera and atmosphere properly. Then use negative prompting for the specific failure modes you’re actually seeing.
The 80/20 rule
Most of your prompt should describe what you want. The negative prompt should handle the relatively small number of things you specifically don’t want.
Positive: A cinematic photograph of an old observatory on a remote mountain at blue hour, wide composition, observatory positioned on the left third, distant mountains fading into atmospheric haze, subtle interior light visible through the windows, realistic stone and weathered metal, 35mm photography, restrained colour palette.
Negative: people, vehicles, text, watermark, excessive HDR, oversaturation, artificial lens flare, distorted architecture
The positive prompt carries the creative direction. The negative prompt protects it. That’s a healthy relationship between the two, and it’s worth noticing how short the negative side stays relative to the positive one.
A reference by category
If you generate regularly, it helps to have a mental (or literal) shortlist by category, so you’re not reinventing the same constraints every time:
Anatomy — extra fingers, fused fingers, malformed hands, duplicated limbs, distorted facial features, asymmetrical eyes, unnatural teeth.
Text and logos — unwanted text, watermark, logo, captions, subtitles, signage. (One exception worth flagging: if the design actually needs readable text, don’t blindly paste “no text” in from a reused negative prompt. A constraint copied from another project can quietly sabotage the one you’re working on now.)
Photography artifacts — excessive HDR, oversharpening, artificial lens flare, excessive bloom.
Realism — plastic skin, artificial reflections, perfect symmetry, CGI appearance.
Composition — centred composition, excessive cropping, subject filling entire frame.
Architecture — warped perspective, impossible geometry, duplicated windows, floating structures.
Products — for something like a luxury watch on stone, “duplicated watch, warped watch face, distorted hands, incorrect numerals, floating object, inconsistent reflections, excessive glare, text, watermark” protects the specific geometry a viewer will actually notice going wrong — quite different from a vague “bad product, low quality.”
Treat these as a toolbox, not a checklist you run through every time. You don’t need every category in every prompt — only the ones addressing a problem you’re actually having.
Sometimes the better move is a positive rewrite
A negative instruction and a precise positive instruction can solve the same problem, and the positive one is usually stronger. Instead of “no centred composition,” try “subject positioned in the far left third, large area of empty negative space on the right.” Instead of “no harsh lighting,” try “soft diffused side lighting with gradual shadow falloff.” Instead of “no plastic skin,” try “natural skin texture with visible pores and subtle tonal variation.” Whenever you can, replace a prohibition with a precise visual instruction, and save the negative prompt for whatever’s left over that genuinely needs excluding rather than redirecting.
Iterate instead of starting with a giant list
Rather than opening with a sprawling negative prompt, work in a loop: generate using only the main creative direction, look at the result, identify the single biggest failure — maybe the hands are wrong, maybe it’s oversaturated, maybe the composition centred itself when you wanted asymmetry — add one or two targeted constraints for that specific problem, and generate again. Over a few rounds, your negative prompt becomes a record of your image’s actual failure modes, which is far more useful than importing someone else’s forty-word list wholesale.
When negative prompts make things worse
If a result starts feeling too sterile, too predictable, too symmetrical, too polished, too generic, or just visually lifeless, the negative prompt may be part of the problem. Creativity needs a degree of freedom, and eliminating every unusual possibility can eliminate the interesting one along with it. Sometimes the strange thing you were trying to remove was the thing that made the image worth looking at — you closed every wrong door, and it turns out the interesting room was behind one of them.
Before adding a negative constraint, run it through three quick questions: is this actually happening, or am I pre-empting a problem I haven’t seen yet? Does it meaningfully damage the image, or is it a minor thing not worth reacting to? And could I describe the desired result more clearly in the positive prompt instead? That last question especially is worth asking before reaching for another exclusion.
A practical two-layer template
Think of a prompt as two layers: creative direction (subject, environment, composition, camera, lighting, materials, atmosphere, style) and constraints (specific unwanted objects, artifacts, style contamination, technical failures).
Creative direction: A lone astronaut standing inside an enormous abandoned observatory, circular architecture disappearing into darkness above, subject positioned in the lower-left third, vast negative space overhead, cold blue ambient light with a single warm practical light behind the astronaut, weathered metal and dusty glass, cinematic 35mm photography, subtle atmospheric haze.
Constraints: text, watermark, logo, extra people, duplicated objects, distorted anatomy, excessive bloom, oversaturation, plastic materials, artificial lens flare
That’s enough. You don’t need another 150 words explaining everything the machine is forbidden from doing.
The strange power of leaving things alone
There’s a temptation to keep refining — one more exclusion, one more safeguard, one more technical instruction — until the prompt becomes a cage, and the generator starts producing technically obedient images that have nothing interesting to say. Negative prompting works best when it creates boundaries, not walls. You want the model to know where the edges are. You still want it to move.
Final checklist
Before generating, ask: is the unwanted element actually appearing, or am I adding it just in case? Could I describe the desired result more clearly in the positive prompt instead? Is my negative prompt specific — have I swapped “bad” and “ugly” for observable visual problems? Am I repeating the same constraint in different words? Am I accidentally contradicting the positive prompt? Is the result starting to feel sterile, in which case I should simplify? And am I pasting in someone else’s negative prompt just because it’s floating around online, or does this one actually belong to this image?
The KROMALOCA Way
A good prompt doesn’t control every pixel. It creates enough direction for the machine to understand the destination, and enough freedom for something unexpected to happen along the way. Negative prompting is part of that — but it isn’t about telling the machine “don’t be weird.”
Sometimes weird is exactly what you came for.
The better instruction is: be weird here, just don’t break this. That’s the difference between suppressing the machine and directing it.
A quick starting formula
When you need somewhere to begin: unwanted objects + unwanted artifacts + unwanted visual behaviour + unwanted style contamination.
unwanted text, watermark, duplicated objects, distorted anatomy, excessive smoothing, plastic materials, oversaturation, artificial reflections, excessive bloom
Then remove anything that doesn’t actually apply to your image. The final negative prompt should be short enough to understand, specific enough to matter, and loose enough to leave the image still alive. The goal was never the safest possible image. It’s the image you actually meant — and if the machine takes a wrong turn, the fix usually isn’t another warning. It’s a better map.