To get a general overview of the UI, please read the Quick Start Guide first.
Training an embedding requires a few different settings than Fine-Tuning or LoRA training. You can enable Embedding training in the dropdown in the top right corner.
To train an embedding, you need to use special prompts. Each prompt needs to include <embedding>
in the place where
you want to place your trained embedding. For example, if you want to train a style of a painting, your prompt could
be a painting in the style of <embedding>
. If you don't want to add a custom prompt for every training image, you can
select "From single text file" as the prompt source of your concept. Then select a text file containing one prompt per
line. An example of such a file can be found in the embedding_templates directory.
If you select "Embedding" as your training method, a new tab called "embedding" will appear. Here you can specify:
- Base embedding: An already trained embedding you want to continue training on. Leave this blank to train a new embedding
- Token count: The number of tokens to train. A higher token count is better at learning things, but will also take up more of the available tokens in each prompt when generating an image.
- Initial embedding text: The text to use when initializing a new embedding. Choosing a good text can significantly speed up training.