Negative Prompts Explained: What to Leave Out and Why

Updated: 2026-11-03 · 6 min read

Sometimes the fastest way to a better AI image isn't adding more words. It's telling the model what you don't want to see.

What a negative prompt actually does

When you write a normal prompt, the image model is pushed towards images that match your words. A negative prompt pushes it away from certain concepts at the same time. Technically, in diffusion models, the negative text replaces the "empty" guidance the model normally compares against, so each step of the generation is steered away from what you listed.

In plain terms: the positive prompt says "go here", the negative prompt says "and stay away from there".

This is why negative prompts are especially useful for fixing recurring problems: blur, watermarks, deformed hands, unwanted text, or a style you keep getting by accident.

Which tools support negative prompts

Support varies, and tools change their features often, so treat this as a general guide:

• Stable Diffusion interfaces (Automatic1111, ComfyUI, many web apps built on SD) usually have a dedicated "Negative prompt" box.
• Midjourney uses the --no parameter at the end of a prompt, for example --no text, people.
• Leonardo and similar platforms often expose a negative prompt field in advanced settings.
• ChatGPT and Gemini image generation don't have a separate box. Instead, describe exclusions inside your request: "no text anywhere in the image", or ask in a follow-up: "remove the people in the background".

A starter negative prompt list

You don't need a giant list. Start small and add only what fixes a problem you've actually seen.

For photographs:

blurry, out of focus, low resolution, overexposed, underexposed, watermark, text, logo, jpeg artifacts

For people:

extra fingers, missing fingers, deformed hands, distorted face, crossed eyes, extra limbs, plastic skin

For clean product or design shots:

cluttered background, shadows on background, reflections, text, watermark, frame, border

For illustrations where you want no photo look:

photorealistic, photograph, 3d render

The mistakes that make negative prompts backfire

Negative prompts are powerful, which also means they can damage an image.

• Overly long lists. Pasting 80 words of "bad anatomy, worst quality, ugly..." can flatten the image, reduce variety and even fight your positive prompt. Many modern models need far less.
• Contradicting your main prompt. If your prompt says "rainy street" and your negative says "water", the model gets mixed signals.
• Negating things that weren't there. Writing "no cats" in a normal prompt sometimes *adds* cats, because the word "cats" appears. Use the proper negative field or --no instead of writing "no X" inside the positive prompt in SD-style tools.
• Using negatives instead of describing what you want. "Not dark" is weaker than "bright soft daylight".

A good workflow

1. Write your positive prompt carefully first: subject, setting, style, lighting, camera, mood.
2. Generate a few images without a negative prompt.
3. Note the specific problems that repeat. Is it blur? Text? A second person you didn't ask for?
4. Add only those words to the negative prompt.
5. Regenerate and compare. If quality drops, remove items one at a time.

This approach keeps your negative list short and purposeful.

Real examples

Problem: a product photo of a perfume bottle keeps showing random text on the label.

Positive: a glass perfume bottle on a smooth beige stone, soft studio light, minimal, centered, product photography

Negative: text, letters, logo, watermark

Problem: a fantasy landscape keeps adding people.

Midjourney: misty mountain valley with a glowing river at dusk, painterly fantasy art --no people, buildings

Problem in ChatGPT: the image includes a caption you didn't want.

Follow-up: "Please regenerate the same scene with no text or caption anywhere."

Negative prompts vs better positive prompts

A negative prompt is a repair tool. The main quality still comes from a clear positive prompt. If you find yourself writing very long negatives, step back and improve the description: name the lighting, the lens, the medium and the mood. Most "ugly" outputs are really "vague" outputs.

Save what works

When a combination finally gives you clean results, save both parts: the positive prompt and the negative list. Over time you'll have a small set of negatives for each type of image: portraits, products, landscapes.

The AI Image Prompts app on Android gives you a categorised library of positive prompts you can copy in one tap and save as favourites. Pair those with the short negative lists above for the tool you use. The app is a prompt collection; you generate the images on your own chosen platform.

Negative prompt templates by use case

Keep a few short templates and pick one per project instead of using a single giant list:

• Portrait photo: blurry, plastic skin, distorted face, extra fingers, watermark – keeps faces natural without flattening the image.
• Landscape: people, buildings, text, oversaturated, haze – useful when you want empty, clean nature scenes.
• Flat illustration or logo: photorealistic, gradient, shadow, 3d, texture, text – stops the model drifting into realistic rendering.
• Food photo: plastic look, burnt, messy plate, harsh flash, hands – keeps dishes appetising.
• Anime character: realistic, photograph, extra limbs, low detail eyes – pushes the style towards clean anime art.

Treat each template as a starting point. If an item is not fixing a visible problem, remove it. A short, targeted list usually beats a long generic one.

Weights and strength

Some interfaces let you weight negatives, for example (watermark:1.3) in Stable Diffusion tools, or by repeating --no terms in Midjourney. Use stronger weights only for stubborn problems, and lower them if the image starts looking dull or broken. Heavy weights can remove detail you actually wanted, so change one weight at a time and compare results side by side.

Testing a negative prompt properly

The only reliable way to know whether a negative term helps is a controlled comparison. Keep the positive prompt, seed (if your tool exposes one), aspect ratio and model version identical, then generate two batches: one with the negative prompt and one without. Look at the same details in both, such as hands, background clutter or unwanted text, rather than judging overall "vibes".

• Change one term at a time. If you add five terms and the image improves, you won't know which one did the work, and the other four may be quietly hurting colour or detail.
• Watch for side effects. Removing "blur" can also remove pleasant depth of field. Removing "dark" can flatten moody lighting you actually wanted.
• Re-test after model updates. A list that worked on one model version can behave differently on the next, so recheck your saved lists when the tool announces an update.

Keep a short note next to each saved list: the tool, the date you tested it, and what it fixed. That note is what turns a random list of words into a reusable tool.

FAQ

Does ChatGPT support negative prompts?

There's no separate negative prompt box. Write exclusions in plain language ("no text, no people in the background") or ask for a revision after the first image.

What is the Midjourney equivalent of a negative prompt?

The --no parameter, added after your prompt, such as --no text, watermark. It works as a negative weight on those words.

Should I always use a negative prompt?

No. Start without one and add negatives only for problems you can see. Unnecessary negatives can reduce quality and variety.

Why did writing "no hat" add a hat?

In many models, mentioning a word in the positive prompt adds that concept, even with "no" in front of it. Put "hat" in the negative field or use --no hat instead.

Do negative prompts fix hands?

They can reduce problems, but they're not a guarantee. Better results usually come from simpler poses, hands not in focus, or newer models, plus inpainting to fix small errors.