Have you been delving into the world of AI art creation? AI text-to-image generation models like Craiyon, Midjourney, and Stable Diffusion have given virtual art a futuristic upgrade, which is why so many people are now using them each day to take their creations to other dimensions.
Released just last year, Stable Diffusion is one of the most popular GAN (generative adversarial network) models. Its deep learning AI has been basically used to generate very detailed images from user text prompts, descriptions that specify what you want the AI to create for you.
However, are you aware that using negative prompts can actually help you to receive more enhanced images? In this blog, we’ll look at Stable Diffusion negative prompts to find out what they are and how they can boost your creative AI output.
Stable Diffusion Negative Prompts — What Are They?
Stable Diffusion, which is a text-to-image, as well as image-to-image deep learning model, works by taking your text prompt to generate images created with artificial intelligence.
And with the release of Stable Diffusion v2, negative prompts have become a thing, because of the way that they allow the user to refine the images generated by the AI service.
Negative prompts technically work as a parameter that instructs Stable Diffusion to get rid of the things that you don’t want to see in your images through the text that you input. A negative prompt can also help the AI to fix abnormalities, and stop it from generating specific things.
Negative prompts technically works as a parameter that instructs Stable Diffusion to get rid of the things that you don’t want to see in your images through the text that you input. A negative prompt can also help the AI to fix abnormalities, and stop it from generating specific things.
So basically, whenever you input a (positive) prompt into Stable Diffusion, that means it will focus on creating images that are associated by the text that you give it. Negative prompts, on the other hand, take away that focus, which means Stable Diffusion will be able to generate more of what you really want.
Negative prompts can also be used to refine your generated images by adding specific parameters like ‘gross proportions’, ’pixelated’, ‘poor lighting’, or ‘unnatural colors’.
This gives you the ability to be more precise when it comes to creating your prompts. In addition, it helps you to receive generated images from the AI that have much more accuracy and detail.
Now you may be wondering if negative prompting is really a necessary step that you need to take, since this is supposed to be artificial intelligence that you’re playing with.
Well, although it’s not absolutely necessary to do so, using negative prompts will help Stable Diffusion to work better for you. That’s because the AI is able to understand them much better than it does the languages that we speak.
If you want an AI to converse with you in this way, then you should go have a conversation with an AI chatbot like ChatGPT instead.
Stable Diffusion Negative Prompts List
When it comes to choosing the negative prompts that you want to use for Stable Diffusion, there is no right or wrong answer for getting the AI to do exactly what you want it to.
Basically, you will have to try and test so that you can learn how the AI works best for you. What you shouldn’t do is simply fill up your prompt with ‘negative’ word vomit, because it will inhibit Stable Diffusion’s ability to create the quality image that you are seeking to produce.
Basically, you will have to try and test so that you can learn how the AI works best for you. What you shouldn’t do is simply fill up your prompt with ‘negative’ word vomit, because it will inhibit Stable Diffusion’s ability to create the quality image that you are seeking to produce.
Having said that, the following is a list of negative prompts that you can use with Stable Diffusion. Using these particular words, or combinations, will help the AI text-to-image generator remove certain parts of the image that you don’t want to see:
- bad anatomy
- bad proportions
- blurry
- boring
- cloned face
- cropped
- deformed
- dehydrated
- disfigured
- dull
- duplicate
- error
- extra arms
- extra fingers
- extra legs
- extra limbs
- fused foot
- fused fingers
- gross proportions
- jpeg artifacts
- long neck
- low quality
- low res
- malformed limbs
- missing arms
- missing legs
- morbid
- mutated hands
- mutation
- mutilated
- obscure
- out of frame
- pixelated
- poorly drawn hands
- poorly rendered face
- poor lighting
- signature
- text
- too many fingers
- ugly
- unclear
- unnatural colors
- username
- watermark
- worst quality
How To Add A Negative Prompt
Now that you have and idea, or know exactly what negative prompt you want to use, all you need to do is add it to your text. This will assist Stable Diffusion to create the image that you want it to.
We asked the AI model to generate an image of the English countryside in the springtime.
Here is our original prompt:
springtime in England, beautiful, atmosphere, vibe, mist, smoke, fire, chimney, puddles, bluebells, hedgerows, creek, lush, bridge, forest, hills, wildflowers, by John Constable, William Hogarth, Joseph Mallord William Turner
And this is what we received:
The most obvious thing to do now is to ask Stable Diffusion to remove anything that we don’t like in the image that it generated of the English countryside. We weren’t keen on the number of people in the bottom right image, so we gave a negative prompt of people:1.5, which told Stable Diffusion that this criteria was now 50% more important.
We also didn’t want Stable Diffusion to give us images that weren’t very distinguishable from the others, so asked it to reduce duplication by adding the ‘no duplicates’ negative prompt:
springtime in England, beautiful, atmosphere, vibe, mist, smoke, fire, chimney, puddles, bluebells, hedgerows, creek, lush, bridge, forest, hills, wildflowers, by John Constable, William Hogarth, Joseph Mallord William Turner, no duplicates, people:1.5
After adding people:1.5 and no duplicates as our negative prompts, this is what we got from Stable Diffusion:
And as you can see, there are a lot more people now and less duplication in our English countryside Constable-inspired images.
For our next example, we decided to get Stable Diffusion to generate a portrait of a woman with red hair wearing a forest green silk blouse who is sitting in an ethereal garden. This was our prompt:
portrait, woman with red hair wearing a forest green silk blouse, sitting in an ethereal garden, marble statue, shining teal river background, futuristic and highly detailed buildings, lifelike, photorealistic, painting, long shot, studio lighting, octane render, by Gerardo Dottori
And this was the result:
OK, so now it was time to refine the generated output and add a negative prompt to our original text:
portrait, woman with red hair wearing a forest green silk blouse, sitting in an ethereal garden, marble statue, shining teal river background, futuristic and highly detailed buildings, lifelike, photorealistic, painting, long shot, studio lighting, octane render, by Gerardo Dottori, poorly rendered face
And here’s what we got by simply adding poorly rendered face as a negative prompt:
However, we still weren’t happy with the hands and fingers situation of our red haired lady, so we decided to add even more negative prompts to our text to see what would happen:
portrait, woman with red hair wearing a forest green silk blouse, sitting in an ethereal garden, marble statue, shining teal river background, futuristic and highly detailed buildings, lifelike, photorealistic, painting, long shot, studio lighting, octane render, by Gerardo Dottori, poorly rendered face, too many fingers, poor lighting, gross proportions
And this was Stable Diffusion’s final result:
So, it’s pretty clear to see that simply by adding negative prompts to the text you input into Stable Diffusion, it will do its best to give you the image you want it to.
Truth be told, most of our AI-image making time is spent with Midjourney, which is significantly better than Stable Diffusion at many things. Photorealistic images in Midjourney mostly look like true photos, not weird slightly abstract artworks (which are cool… they’re just not photorealistic!). The header image in this article is a midjourney image with the prompt simply “a negative prompt”.
Cool image.
Midjourney is a hassle to setup, you have to use Discord, and you have to pay to use it a significant amount. But it’s so worth it if you really enjoy working with AI art.
Conclusion
Being able and knowing how to use negative prompts is essential when using an AI text-to-image generation model like Stable Diffusion.
So, what are they exactly? Well, Stable Diffusion negative prompts basically work by acting like a guide so that the AI doesn’t stray too far away from the subject matter at hand.
We hope that this blog post has given you some idea on how using Stable Diffusion negative prompts can help you to refine and enhance the generated images that you get from the AI
