What Is a Negative Prompt
Stable Diffusion has two boxes. One is the normal prompt. The other is the negative prompt.
The normal prompt tells the model what you want. The negative prompt tells the model what you do not want.
When I started, I only used the first box. My images had extra fingers and blurry faces. Then I learned about negative prompts. My results got much better.
So this is simple idea. You push away bad stuff. The model then focus more on good stuff.
Why Negative Prompts Matter
Stable Diffusion is not perfect. It make mistakes. Common mistakes are:
- Weird hands and too many fingers
- Blurry or grainy photos
- Extra arms or legs
- Ugly faces
- Text or watermarks you never asked for
Negative prompt is your tool to fix these. You do not need to change your main prompt. You just add words to the negative box.
In my experience, this save a lot of time. Before, I would generate many images and throw most away. Now I get more usable images in less tries.
My Basic Negative Prompt List
I keep a list that I paste into almost every image. Here is what I use:
blurry, low quality, bad anatomy, extra fingers, extra limbs, deformed, ugly, watermark, text, signature, jpeg artifacts, grainy
This is good starting point. It removes most common problems. You can copy this and use it right now.
For people and portraits, I add more:
bad hands, missing fingers, fused fingers, mutated hands, poorly drawn face, asymmetric eyes, cross-eyed
Hands are the hardest part. The model struggle with them a lot. So I put many hand words in negative box.
Match Negative Words to Your Image Type
One thing I learn is you should not use same negative prompt for everything.
If you make a photo, add words like cartoon, anime, painting, drawing, 3d render. This push the model toward real photo look.
If you make anime art, do the opposite. Add realistic, photo, photograph to negative box. This keep the style clean and drawn.
For landscapes, you do not need hand words. Instead use people, person, blurry, distorted horizon.
So think about what you want. Then think about what would ruin it. Put the ruin words in negative box.
Do Not Add Too Much
When I first learn this, I got excited. I put fifty words in negative box. I thought more is better.
It was not better. My images became flat and boring. The model was too scared to make anything.
So now I keep it shorter. Maybe ten to fifteen words. This is enough for most images.
If you add a word and the image gets worse, remove it. Test one change at a time. This is the best way to learn what works.
Use Weights for Strong Control
Some tools like Automatic1111 let you add weight to words. You do this with brackets.
For example: (blurry:1.4). This make the word stronger. The model try harder to avoid blurry.
Be careful with high numbers. If you go too high, like (blurry:2.0), it can cause strange results. I usually stay between 1.1 and 1.5.
This is small trick but very helpful. Use it when one problem keep coming back.
A Simple Workflow I Follow
Here is how I work step by step:
1. Write my main prompt first.
2. Generate one image with empty negative box.
3. Look at what is wrong.
4. Add negative words for those problems only.
5. Generate again and compare.
This way I understand what each word does. I do not just copy random lists from internet.
Over time, you build your own list. You know which words help your style. This is real skill and it come from practice.
Final Thoughts
Negative prompts are powerful but simple. You just tell the model what to avoid.
Start with a basic list. Change it based on your image type. Do not add too many words. Use weights when you need more control.
I found that good negative prompt is half the work. The main prompt gives the idea. The negative prompt cleans the mess.
Try it on your next image. Add a few words and see the difference. Small change, big result.