How to Use Negative Prompts and Fix Hands?

In this guide, you will find methods on how to fix hands in stable diffusion and generate natural-looking eyes. Additionally, you will learn more about stable diffusion negative prompts and find a list of prompts for you to use every time you generate an image.

What is stable diffusion?

Stable Diffusion is a groundbreaking tool that uses deep learning to generate high-quality images from text prompts. Developed by Stability AI, it’s one of the most advanced diffusion models currently available due to its ability to produce highly realistic and coherent outputs. When provided with a text description, Stable Diffusion utilizes its vast knowledge base to conceptualize the prompt before rendering the requested content. Through a technique known as diffusion, it can smoothly transition from a random noise pattern to the target output.

As of writing, the latest stable diffusion models are Stable Diffusion 2.1 and Stable Diffusion xl 1.0. SDXL 1.0 is a sort of lite version of the image-generating tool. Both models demonstrate an impressive understanding of concepts, objects, artistic styles, and more. Stable Diffusion has the potential to revolutionize content creation across industries by unleashing human-level imagination through simple text inputs.

Negative Prompts:

You might be asking, what is a negative prompt? Negative prompts allow you to avoid unwanted results by describing what you don’t want the AI to generate. This helps steer the output in a more desirable direction. Common things to exclude with negative prompts include explicit or objectionable content, logos/brands, text/writing, certain colors, types of scenes, etc.

Below is a list of good negative prompts for Stable Diffusion.

Note: It’s best to experiment with negative prompts. According to your art style, you might not need to use every single prompt listed here.

Universal Negative Prompts:

There are some universal negative prompts for Stable Diffusion that you might find useful enough to add to every case.

Signature	Tiling	Caricature
Watermark	Low Contrast	Body Horror
Username	Underexposed	Lowers
Error	Overexposed	Worst Quality
Text	Bad Art	Low Quality
Jpeg Artifacts	Beginner	Duplicate
3d Max	Amateur	Morbid
Cropped	Dehydrated	Mutilated
Cut Off	Blur	Mutant
Out Of Frame	Blurry	Mutation
Body Out Of Frame	Haze	Mutated Hands
Distorted Face	Missing Arms	Poorly Drawn Face
Extra Fingers	Missing Legs	Deformed
Extra Limbs	Cloned Face	Bad Anatomy
Extra Arms	Poorly Drawn Hands	Bad Proportions
Extra Legs	Poorly Drawn Feet	Disfigured
Gross Proportions	Fused Fingers	Long Neck
Malformed Limbs	Too Many Fingers	Disfigured Eyes
Desaturated	Ugly	Distortion

Note: Even though I’ve added prompts like ‘bad anatomy’ and ‘bad proportions’, it might not do anything. This is because the AI doesn’t know what a ‘good’ anatomy is. These prompts are added only as a countermeasure and are not a definitive way to fix anything.

Stable Diffusion Fix hands:

Using the above negative prompts should be able to fix the hands and fix eyes in most cases. The AI might not know what a good anatomy is, but we can be sure that it knows how many fingers there are on a human hand. If you still can’t get the hands right even after using these negative prompts, there are a few other things you can try.

Try adding some positive prompts:

Experiment with some of these prompts. Some users have reported these to work perfectly for them.

• beautiful hands
• intricate hands
• transparent nail-polish
• artist study hands
• anatomically correct
• Leonardo da Vinci

Use Embeddings:

Embeddings use a technique known as textual inversion, users can define new keywords or concepts that the model can respond to without needing to retrain its parameters. Textual inversion works by finding the closest matching embedding vector to represent the new term being introduced. It essentially expands the internal vocabulary of the AI without altering the underlying model architecture. By utilizing textual inversion of embedding vectors, prompts can influence generations in novel ways even with terms the AI was not explicitly trained on.

You can download these embeddings from civitai. Try some of these:

• Bad-Hands-5
• negative_hand
• better hands
• badhandv4
• Eye – LoRa
• beautiful detailed eyes
• Fast Negative Embedding
• EasyNegative

Embeddings are made to target a specific style or aspect of an image. Adding embeddings to your prompts is almost always better than simply using a long list of negative prompts. Less is better.

Inpainting:

If nothing else works, img2img inpainting would be your best bet. Inpainting allows you to mask a certain part of your image and redo the prompts on that specific part only.

• Send your generated image to Inpaint – you should see an option to do that in the im2img tab
• Carefully draw over the hand or eyes – whichever you want to fix
• Add prompts for the specific elements you are trying to fix

You might have to try this a few times to get the results you want. Another way to get natural hands is:

• Search Google images for the image of a hand that most matches the one you generated
• Use an image editing program like Photoshop to rotate the image and match the lighting
• Crop the image and place it over the broken hands
• Now upload this image to Inpaint and let the AI fix the image

Depth Map:

If you are using AUTOMATIC1111’s stable diffusion webui, There is an extension you can use for the hands. It’s called Depth Map. It lets you control the shape of the hands with sliders. It is worth checking out as it is made specifically to generate proper-looking hands.

Afterword:

There is no surefire way to fix the anatomy of an AI-generated image. Negative prompts that you used in one image might not work for another. This is because the image prompts and the negative prompts heavily influence each other.

Experimenting with negative prompts takes some trial and error, but it’s rewarding to see your conceptual ideas come to life. And while negative prompts provide control, don’t forget to also embrace serendipity. I wish you the very best in your artistic endeavors. Happy image-making!