Negative embeddings

What is embedding?

Embedding is a method used to introduce new keywords or concepts into a model without modifying the model itself. A new keyword is defined for the desired object or style that is not already present in the model. This new keyword is tokenized. the token is then converted into a unique embedding vector, which is a numerical representation used by the model for image generation.

Overall, embedding offers a way to expand the capabilities of a model by introducing new keywords and associated embedding vector, thereby enabling the generation of images or outputs in different styles or with new objects.

What is negative embedding?

This is an embedding used in the negative field to avoid styles or objects that you don't want to see, such as a poorly-formed hand. Negative embedding is actually trained using images of undesirable contents. For instance, the negative hands embedding is trained using images of hands with missing fingers, extra fingers, or deformed shapes, with the aim of creating an embedding that can represent undesirable hands. When this negative embedding is employed in the negative field, it introduces the concept of bad hands to the model and attempts to avoid generating them.

How to use it?

Click on "Negative embeddings" in creation page.

If you are unfamiliar with all these negative embedding models, you can click on the model's name. This will take you to the model detail page, where you can find the model's introduction and view example images.