References enable you and your team maintain a consistent visual language across generations. Once created, a Reference is accessible to anyone within the workspace.

Create New Reference

There are three ways to create a Reference:

1. References page: After you login, go to the Reference page and click on New Reference

2. Composer: From the composer, click the +, and then + again to create a new Reference

3. Canvas: Select an image or group of images and click + Reference in the sidebar.

A reference can contain images, a prompt, or both. Notice in the prompt field the use of the {prompt} token. This syntax indicates where the user prompt should be inserted. This is an optional feature — by default the user prompt will be inserted at the end of the style prompt.

Best Practices

• Separate style from subject: References should focus on style information and avoid mentioning people, objects, and locations. This enables them to be used across a wide range of subject matter.

• Put the most important words first. Words at the beginning of the prompt are weighted more heavily than words at the end of the prompt.

• Experiment with prompt placement: To control where the user-provided prompt is placed, use {prompt} in your style prompt. By default the user prompt will be inserted at the end of the style prompt. Placing it at the beginning can produce different results.

Using References

Once you have created a Reference, you will be able to choose it from the Reference Picker in the composer.

Trained References

Trained References are References that been fine-tuned on a custom dataset. Training a Reference can take about 5 minutes to complete and costs 100 volts, and can produce more consistent results than prompting alone. The most important thing to consider when training a Reference is the quality of your data.

To train a Reference, hover over the Reference card, click on the ••• and then click on Train.

Data Collection Tips

• Maintain a consistent style: Make sure the style is consistent across your dataset, otherwise the model will not produce great results.

• Use diverse subject matter: Use images that show a wide range of people, objects, and locations. Data sets that have the same subject matter can become overfit on those examples.

• High resolution images: 1024×1024 resolution or greater is highly recommended.

Here's an example of what a great dataset looks like: