What is a negative prompt?

It's a separate piece of text, alongside your main prompt, that tells an image model what to steer away from, such as blurry or extra limbs. It only works on models with a true second conditioning channel, like Stable Diffusion tools or Midjourney's --no parameter.

Do all AI image models support negative prompts?

No. Many of today's natural-language, instruction-following image models read one single prompt and have no separate way to subtract a concept, so writing exclusions into them does little or nothing.

Why did the thing I tried to exclude show up anyway?

On single-prompt models, the exclusion word is still read as text by the model, so naming a thing to avoid can sometimes bring it into the image. Describing what you want instead, rather than what you don't, avoids this.

What's a typical negative prompt word list?

Common entries include blurry, low quality, watermark, extra fingers, bad anatomy, duplicate, and oversaturated — each targeting a specific, well-known failure mode in diffusion output.

How do I write a negative prompt in Plurel?

Plurel's Image tool uses one prompt box for every model, so instead of a negative field you write a positive description of exactly what you want and fix anything left over with the Edit or Paint tools.

Can I remove an artifact after the image is already made?

Yes. Send the image to Plurel's Edit tool and describe the fix as an instruction, or use Paint to mask just that region and regenerate only that part.

Negative Prompt Guide: What to Leave Out of AI Images

What a negative prompt actually is

A negative prompt is a second piece of text you give an image model alongside your main prompt, telling it what to steer away from. Instead of describing the picture you want, you describe the picture you don't want — "blurry, watermark, extra fingers" — and the model pushes its output away from those concepts during generation.

This only works on models built around a technique called classifier-free guidance (background reading here). In plain terms: the model generates the image twice in parallel math, once pulled toward your positive prompt and once pulled toward your negative prompt, then the final image is steered toward the first and away from the second. Stable Diffusion-based tools (Automatic1111, ComfyUI, and most "open" image generators) keep this as a literal second text box. Midjourney exposes it through a --no parameter you add to the end of your prompt line instead of a separate field, a syntax covered in detail elsewhere if you want the full parameter list.

A negative prompt only does something if the model you're using actually has that second conditioning channel — plenty of today's most popular image models don't.

Which models honor a negative prompt and which don't

The split in the market right now is roughly: older or more "classic" diffusion interfaces keep a true negative-prompt channel, while the newer wave of natural-language, instruction-following image models (the kind built on large multimodal transformers) mostly don't have one at all. For those, you write one prompt, and anything phrased as "no X" or "without X" is just more text for the model to read — it can't mathematically subtract it the way classifier-free guidance does, so the safest assumption with any instruction-style model is that there is no reliable exclusion channel and you should write only what you want to see.

This matters for wording: a one-box, natural-language model reads your whole sentence, including the word you were trying to ban. If you type "a kitchen with no clutter on the counter," it has no clean mechanical way to subtract clutter — it's just reading a sentence that contains the word "clutter." The fix across every model type is the same: describe the counter you want (clear, empty, wiped clean) instead of describing the mess you don't want.

Common artifacts and the words that cause (or fix) them

Certain words show up again and again in negative-prompt lists because certain failure modes show up again and again in diffusion output. Here's what each one is actually targeting:

Artifact	What it looks like	Positive fix to write instead
Extra/fused fingers	Six fingers, melted hand	"hands resting flat, fingers spread, visible palm"
Bad anatomy	Twisted limbs, wrong proportions	"anatomically correct, standing straight, full body in frame"
Watermark/signature ghosting	Faint text or logo shapes in a corner	describing a clean scene with no mention of branding at all
Low detail / plastic skin	Waxy, over-smoothed faces	"visible skin texture, soft natural light, 85mm portrait lens"
Duplicate subjects	Two heads, cloned limbs	"a single person, centered, one subject only"
Oversaturation	Neon, blown-out colors	"muted natural colors, soft contrast"
Wrong style bleed	Cartoon lines on a photo request	"photographic, shot on camera, realistic lighting"

Notice the pattern: every "negative" word in that left column has a positive description on the right that does the same job without requiring the model to understand negation at all. That's the technique that works everywhere, regardless of which model you're using.

One quirk worth knowing: typing an exclusion word can sometimes make the exact thing you banned appear, because the model's text encoder still "reads" that token even when it's meant as an instruction to avoid it. This is most common with single-prompt-box models and is the main reason positive phrasing outperforms negative phrasing as these models get more capable.

How to do it in Plurel

Plurel's Image tool uses a single prompt box for every model — there's no separate negative-prompt field, because most of the image models in the app (Seedream 5, Nano Banana 2, GPT Image 2.5, Qwen Image 3, Grok Imagine 2) are the natural-language instruction-following kind, not classic Stable Diffusion UIs. So the technique above is the technique to use in Plurel:

Open the chat at plurel.tech and tap Image in the row under the composer, or go straight to plurel.tech/?model=image.
Pick a model. For a quick test, Plurel Lite (0.5 credits) is cheapest to iterate on wording. For a final photorealistic shot, use Nano Banana 2 (4 credits) or GPT Image 2.5 (3.5 credits).
Write only what you want in the frame — skin texture, pose, lighting, number of subjects, style — and leave out any mention of the artifact you're trying to avoid.
Generate, then look at exactly what went wrong.
Fix it with a follow-up message rather than rewriting the whole prompt. Plurel's chat edits what was just made, so "remove the extra hand on the left" or "make the background plain white" works as a direct reply.
If something unwanted is still stuck in the image after a couple of tries, send it to Edit (same credit cost as the model that made it) and describe the fix as an instruction, or use Paint to mask just that region and regenerate only that part.

If you're asking one of the chat models — Claude Opus 5.5, Claude Sonnet 5.5, or GPT-6 Astra — to plan and make the image for you, you can literally tell it your exclusions in plain English ("make sure there's no text in the image, only one subject, hands not visible") and it will write the positive-phrased prompt for whichever image model it picks, then show you the plan and credit cost before generating.

Ten prompts: exclusions rewritten as instructions

Each of these replaces a "don't" with a "do." Copy them as-is.

A close-up studio portrait of a woman's face, soft directional light from the left, natural skin texture, no visible hands in frame, shallow depth of field

Produces a clean headshot without the hand-artifact risk entirely, by keeping hands out of frame. Run on Nano Banana 2, 4 credits.

A single golden retriever sitting centered on a wooden porch, sharp focus, crisp natural daylight, high detail fur texture

Avoids duplicated or blurry subjects by specifying "single" and "centered." Run on Plurel Pro, 1.5 credits.

A minimalist white ceramic mug on a plain grey background, studio lighting, product photography, no props, no branding visible

Keeps the shot clean for a product listing. Run on GPT Image 2.5, 3.5 credits.

A medieval stone village street at dawn, torches and wooden carts, muted earth-tone palette, every building period-accurate

"Period-accurate" does the work that "no modern objects" would try and fail to do. Run on Seedream 5, 3 credits.

A hand holding a coffee cup, palm and fingers fully visible and relaxed, resting on a table, natural proportions

Describing the exact hand position sidesteps the extra-finger failure mode directly. Run on Nano Banana 2, 4 credits.

A tidy modern kitchen counter, empty and wiped clean, one bowl of fruit as the only object, bright even lighting

"Empty and wiped clean" replaces "no clutter." Run on Plurel Lite, 0.5 credits.

A storefront sign reading "OPEN TODAY" in bold sans-serif letters, crisp and legible, centered on a red awning

Spelling out the exact wanted text avoids the garbled-text problem instead of trying to ban bad text. Run on Qwen Image 3 Pro, 2.5 credits.

A single red sports car on an empty asphalt road, muted natural sky, realistic reflections, photographic lighting

"Empty asphalt road" and "single" prevent duplicate cars and crowded backgrounds. Run on Grok Imagine 2, 2 credits.

A phone wallpaper of a calm ocean horizon at sunset, soft gradient sky, no objects in the water, vertical 9:16 composition

Simple, empty composition described directly rather than listing what to remove. Run on Plurel Pro, 1.5 credits via the phone wallpaper tool.

A realistic photo of an office desk with a laptop, notebook, and pen only, neutral wood surface, soft window light

Naming the exact three objects stops the model from adding random extras. Run on Seedream 5, 3 credits.

Troubleshooting

The artifact keeps coming back no matter how I phrase it. Generate first, then fix the specific region with Paint rather than fighting the whole prompt — masking just the problem area and regenerating it is faster than rewording five times.

I wrote "no text" and text appeared anyway. This is the known quirk above — drop the word entirely instead of negating it, and if you need the surface clean, just don't mention anything that lives on it.

My Stable Diffusion-style negative prompt list doesn't translate. That's expected outside true negative-prompt tools. Convert your list to positive description using the table above rather than pasting the same negative list into a single-box model.

For more on why models fail in specific, predictable ways (hands, text, physics), see Why AI Video Looks Wrong, and for the underlying mechanism behind why some models take a negative prompt and others don't, see How AI Image Generators Work.

Write the picture you want, not the one you're avoiding, and most of what a negative prompt used to fix never shows up in the first place.