What Are Negative Prompts?
When you generate an AI image, your text prompt tells the model what to include. A negative prompt does the opposite — it tells the model what to exclude or avoid. Most AI image generation platforms provide a separate input field for negative prompts, and learning to use it effectively is one of the fastest ways to improve your results.
Think of it like giving instructions to a photographer: your positive prompt says "I want a portrait of a woman in dramatic lighting" and your negative prompt says "no harsh shadows on her face, no distracting background elements, no lens flare".
Why Negative Prompts Work
Diffusion models generate images by gradually reducing random noise, guided by both a positive and negative condition. When you provide a negative prompt, the model is steered away from generating features associated with those terms — it literally subtracts that direction from the generation process at each diffusion step.
This makes negative prompts especially useful for:
- Removing common artifacts that appear even in good generations
- Correcting style drift (preventing realistic prompts from going cartoon)
- Fixing anatomical issues like distorted hands or extra fingers
- Cleaning up unwanted background elements
- Improving overall sharpness and detail quality
A Reliable Base Negative Prompt
Most experienced AI artists keep a base negative prompt that they use for almost every generation, then add to it for specific cases. Here is a solid starting point:
blurry, out of focus, low quality, low resolution, pixelated, grainy, noisy, jpeg artifacts, overexposed, underexposed, bad composition, watermark, text, logo, signature, username, cropped, poorly drawn, bad anatomy, extra limbs, missing limbs, deformed, disfigured, ugly, duplicate, morbid, mutilated, extra fingers, fused fingers, mutation
This covers the most common failure modes: quality degradation, watermarks, and anatomical errors. It works across most subjects and styles without overly constraining the generation.
Style-Specific Negative Prompts
For Photorealistic Images
Add these to prevent the image from drifting into illustrated or animated territory:
cartoon, anime, illustrated, painting, drawing, sketch, 3D render, CGI, digital art, unrealistic
For Digital Art and Illustrations
If you want clean, professional illustration work:
photorealistic, photograph, photographic, 3D render, CGI, low detail, flat, boring, generic
For Portrait Photography
Portraits often suffer from specific issues that these terms address:
closed eyes, looking away, unflattering angle, red eye, skin blemishes, heavy makeup (if undesired), multiple people (if a solo portrait is intended)
For Landscape Photography
people, text, buildings (if natural landscape is desired), power lines, cluttered, busy
Common Mistakes When Using Negative Prompts
Over-loading the Negative Prompt
Adding hundreds of terms to your negative prompt can actually hurt results. The model has limited "attention" to distribute across all terms. Very long negative prompts dilute the effect of each individual term. Stick to 20–40 terms that are most relevant to your use case.
Using Contradictory Terms
Avoid creating conflicts between your positive and negative prompts. If you want an impressionist painting but add "painting" to your negative prompt, the model receives contradictory instructions and the result suffers.
Negating Concepts That Are Hard to Remove
Some qualities are difficult to fully remove with negative prompts. For example, trying to remove all shadows from a naturally dramatic lighting setup will produce strange results because shadows are a fundamental part of how that lighting works.
Testing and Refining Your Negative Prompts
The best way to understand how negative prompts affect your specific use case is to experiment systematically. Generate a baseline image, then add one negative term at a time and compare the results. This approach reveals which terms are actually effective for your prompt and which have little effect.
Keep a reference document with your best-performing negative prompt combinations. Over time, you'll develop a toolkit of negative prompts tuned to your preferred styles and subjects.
Pre-Built Negative Prompt Templates by Use Case
Here are tested, ready-to-use negative prompt templates for the most common generation scenarios. Copy, paste, and customise:
Universal Quality Baseline
blurry, out of focus, low quality, low resolution, pixelated, grainy, noisy, jpeg artifacts, watermark, text, logo, signature, cropped, poorly drawn, bad anatomy, deformed, disfigured, extra limbs, missing limbs, mutation, duplicate
Use this as the base for everything. Add the category-specific terms below on top of this.
Photorealistic Portraits
cartoon, anime, illustrated, painting, drawing, sketch, CGI, 3D render, digital art, unrealistic, unflattering angle, bad skin texture, cross-eyed, asymmetrical face, overexposed, plastic skin
Landscape Photography
people, tourists, cars, power lines, buildings (if unwanted), cluttered, lens flare, overexposed sky, flat lighting, desaturated, washed out
Digital Concept Art / Illustration
photograph, photorealistic, CGI render, amateurish, flat colours, inconsistent style, childish, low detail, scribbled
Product Photography
shadows on product, dirty, scratched, worn (unless desired), blurry product, background elements, busy background, cluttered, warped, distorted product shape
Anime / Manga
photorealistic, 3D render, Western cartoon style, poorly drawn faces, inconsistent style, off-model, bad proportions
Understanding Negative Prompt Weight
Some AI platforms let you assign weights to negative prompt terms, similar to positive prompts. The syntax varies by platform and model, but the concept is the same: a term with higher weight has more influence on excluding that concept from the generation.
For terms that are particularly persistent — such as blurriness in styles that naturally tend toward soft focus, or extra fingers in complex hand poses — increasing the weight can help. Start at 1.2–1.5 and see if the issue improves without over-correcting and making the image look sterile.
Negative Prompts for Specific Problem-Solving
Fixing the "Too Many Fingers" Problem
Anatomically correct hands remain one of the hardest challenges for diffusion models. If your portrait subjects consistently have hand issues, add this to your negative prompt:
extra fingers, fused fingers, missing fingers, distorted hands, bad hand anatomy, unrealistic hands
Additionally: if you're generating a portrait where hands aren't the focus, phrase your prompt to not include hands ("close-up portrait", "head and shoulders") so the model doesn't need to render them at all.
Eliminating Unwanted Text Artifacts
Many models inject text artifacts — especially at lower quality settings or on certain subject matter. Prevent it with:
text, writing, letters, numbers, words, captions, subtitles, watermark, signature, label
Preventing Style Contamination
If you're generating in a specific style and finding it keeps drifting toward something else (e.g., your oil painting keeps looking digital, or your photography keeps looking illustrated), aggressively exclude the competing styles:
For oil painting: digital art, vector art, 3D render, photograph, CGI
For photography: painting, illustration, drawing, digital art, cartoon, anime
Frequently Asked Questions About Negative Prompts
Should I always use negative prompts?
For most generations, yes — at minimum the quality baseline. The exception is when you're doing abstract art or experimental generation where "artifacts" might actually serve your creative intent. For professional or commercial work, negative prompts are almost always beneficial.
Do negative prompts slow down generation?
Marginally. They add a small amount of computation to each diffusion step. The quality improvement is almost always worth the negligible time cost.
Why does a term in my negative prompt still appear in the output?
A few reasons: the term may be too generic (the model interprets "bad" differently than you might expect), the concept is deeply associated with the subject matter and difficult to remove, or the term weight is too low to override the positive prompt's pull. Try more specific terms, increase the weight, or use image-to-image with a better base image.
Ready to experiment? Open the ImageGen generator and try adding a negative prompt to your next generation. The improvement is often immediate and significant.
Negative Prompt Starters by Use Case
Here are complete, copy-paste-ready negative prompt starters for the most common generation types. Use these as your baseline and add to them as needed:
Portrait Photography
ugly, disfigured, deformed face, crossed eyes, asymmetrical eyes, bad teeth, extra fingers, mutated hands, poorly drawn hands, blurry, watermark, text, logo, bad anatomy, worst quality, low quality, jpeg artifacts, duplicate, morbid, mutilated, out of frame
Landscape and Nature
people, humans, buildings, urban elements, power lines, text, watermark, logo, blurry, overexposed, underexposed, bad composition, worst quality, low quality, jpeg artifacts
Product Photography
background clutter, shadows, reflections, watermark, text, logo, distorted product, fake looking, plastic sheen, overexposed, colour cast, worst quality, low quality, blurry
Illustration / Concept Art
photograph, realistic, 3D render, CGI, low detail, rough sketch, unfinished, bad proportions, worst quality, low quality, watermark, text, logo, jpeg artifacts
Architecture and Interiors
people, vehicles, bad perspective, distorted geometry, unrealistic proportions, low quality, blurry, overexposed, underexposed, watermark, text
The Weight System Explained
On Stable Diffusion-based platforms, you can control how strongly a negative prompt term is applied using weight syntax: (term:1.5) increases the weight by 1.5× — the model works harder to avoid that element. (term:0.7) reduces the weight — it applies the exclusion more gently.
Most of the time, default weight (no parentheses) is sufficient. Use increased weights when:
- A specific unwanted element keeps appearing despite being in the negative prompt
- A style element is contaminating your desired style strongly (e.g., anime features bleeding into a realistic portrait)
- A quality issue is severe and needs strong suppression
Avoid setting weights above 1.8–2.0 — excessively high weights can paradoxically cause artefacts as the model pushes too hard against the excluded concept, introducing distortion in the process.
Building a Personal Negative Prompt Library
Just as you should maintain a library of effective positive prompts, keep a personal negative prompt library — a set of tested combinations that reliably improve your outputs for specific categories. Over time, you will discover that certain terms solve problems that come up repeatedly in your generations.
A good negative prompt library is organised by:
- Quality baseline: Terms you include in almost every generation (low quality, blurry, watermark, etc.)
- Category-specific: Terms you use only for specific subjects (extra fingers for portraits, plastic sheen for products)
- Style-protection: Terms that prevent unwanted style drift (e.g., "anime" when generating realistic portraits, "photograph" when generating illustrations)
Start with the templates in this article and refine them based on the specific artefacts you see in your own generations. Your personal negative prompt library will be tuned to the specific model and use cases you work with, and will become one of your most valuable creative assets.