Create and train custom models

From the Firefly home page, select Custom Models, and then select Create custom model.

Select the type of model you want to create based on whether you're training for a style or a specific subject. If you're an enterprise user, you can select Start with your own use case to start from scratch.

Select Browse images and upload 10–30 images that represent the style or subject you want the model to learn.

Make sure to choose images with variations you want the model to generate, such as different perspectives, compositions, or backgrounds. This helps the model learn the range of outputs you need.

Once uploaded, the analysis and captioning process should begin immediately. In case analysis doesn't begin, select Analyze and caption.

Review the generated captions for each image and correct any inaccuracies or add details that better describe the image's style or subject.

Verify that each image adheres to best practices and resolve any issues flagged by Firefly before continuing. You can hover over the   icon on each image to find out how you can improve it.

For subject-based models, such as a character, person, or still-life object, make sure to include a Name and a Concept ID. Select the Model properties   icon from the left rail to view and edit these and other properties:

• Name: The name of your custom model 
• Description: A description of the style the model generates
• Sample prompt: A starting prompt that people can fine-tune when generating new content in your style 
• Concept ID: Only for models trained on a specific subject (such as a person, character, or still-life object): a unique name you'll use in all captions and prompts to refer to the specific subject you’ve trained a model on

Review the generated tags for the model by selecting the Model tags   icon from the right rail. Make sure that at least three tags describe permanent characteristics of your style or subject.

Select the Model analysis   icon to open the Model analysis panel and review the quality of your training dataset. Finally, you'll need to review and address any recommendations for improving your model.

If you made changes to captions, tags, or metadata, select Analyze and caption again before proceeding to training.

Select Train to begin the training process.

When prompted, select Agree and Proceed on the Permission to train on images dialog box. This helps you grant necessary permissions to Firefly and confirm that you have the rights to use the uploaded images.

Wait for the training to complete. This may take 30 minutes to several hours, depending on the complexity of your dataset. Once the training is complete, you'll receive an email notifying the same.

After receiving the email notification, select Generate in to test your model with sample prompts.