Negative Embeddings

¶ What is a Negative Embedding?

Negative embeddings are used in the negative prompt field, but they are actually trained models created from images. Bad images. Images of things you want to avoid, like mangled arms and extra fingers.

¶ How do I use them?

The negative embeddings we’ve added to MonAI are all listed in a separate section under the suggested negative prompts. That way, you can see which ones are available and just tap on one (or more) to add them to the negative prompt field.

Negative embeddings often have some impact on aesthetics - composition, color - but they are mainly used for human anatomy. If you are trying to generate an animal, you may be better off without a negative embedding.

¶ Which Negative Embedding should I use?

Some negative embeddings are geared towards anime, while others are designed for photorealism. Some are made to go with specific AI models. You can use an anime negative embedding with a photorealistic AI model, but the results MIGHT have more defined lines (like anime) and look less photorealistic. All the negative embeddings seem to work well with anime.

¶ Why do embeddings take up so much character space?

The negative prompt field allows 150 tokens. This is illustrated in characters (0/500). easynegative only uses 8 tokens, but ng_deepnegative_v1_75t uses 75. This means, if you use ng_deepnegative_v1_75t you will be using more of the negative prompt space that the sum of the characters used.

MonAI creates a rough estimate of the amount of characters a negative embedding would use, but these are ESTIMATES ONLY, and they are high estimates. The chart below shows how many tokens each negative embedding takes so that you can compare them to each other.

¶ What’s with all the Unspeakable Horrors?

The v stands for vector. The higher the number, the more powerful it is, and the more tokens it uses. Does that mean it’s better? I’ve gotten better results with 24v and 32v, but 64v is also popular.

The lower vector versions are better for photorealistic images and for AI models that already produce high-quality results. The higher vector versions tend to change the image style significantly. Still, they can achieve outstanding results.