How to train Firefly Custom Models

On the Adobe Firefly homepage, select Custom models.

Browse popular use cases you can train models for. Select one of the use cases to train your own model for that use case. These include:

• Lifestyle photography
• Photoshoot of a person
• Still life photography
• Illustrated character
• Iconography
• Illustrations
• Isometric and 3D graphics
• Explore new brand expression possibilities
• Backgrounds for product shots

If you don’t see the right use case listed, select Start with your own images. 

When you create your own custom model, you can select to train it on a subject, style, or iconography.

Review the model version. You can train custom models on Firefly Image Model 5 (preview). During the preview period, you may choose to train models on the previous version, Firefly Image Model 4, by selecting Image 4 from the right panel of the model editor. Switching models will regenerate captions and model tags, overriding existing content.

Review the Model Score (beta) and its recommendations. Model Score is shown to the left of the Train button. This score (1–100) evaluates your entire training set against best practices for creating custom models. Quality training sets are correlated with quality results.  

• If the score is below 85, select Model Score to open a panel that displays specific suggestions for improving your training set (e.g., replacing images that don't align with the style of other images).
• Apply these improvements before submitting the model for training. 

The score evaluates the entire training set against best practices for creating custom models.

Review the model metadata that we automatically generate for you based on what you upload: 

• Model title: The title you will see when selecting your model. 
• Model description: A description of what the model generates.
• Sample prompt: This gives individuals who use your model a starting point to decide what prompts to use with your model. 
• Model Tags: Permanent attributes of the subject or style you're training a model on, such as brown hair for a brunette character. Do not include changeable attributes in Tags, like what object a character is holding.
• Captions: Notable parts of each training image are described using language similar to what you will prompt with, such as “young woman wearing a green sundress and holding a seashell at the beach, photorealistic.” Models trained on a subject (e.g., character, person, or still-life object) should include the model Concept in each caption. 

Select Train. This may take a few hours. You will see a progress indicator that estimates the amount of training you have completed.

You may close this tab and open your model again from Model Inventory.

Models will typically show the following statuses:

• Draft before training
• Training while training
• Ready once training is finished

It may show cancelled if you choose to cancel its training or if it hits an unexpected error that needs attention before submitting it for training again.

You can test that your model matches your intention before publishing and sharing it.

From the model inventory, hover over a model and select More menu   icon > Preview and Test. Or, open a model and select Preview and Test.

Enter a prompt and use Generate to preview the type of images that the model will generate. Repeat as desired.

To publish them, go to model inventory, hover over a model, and select More Menu > Publish. Or, open a model and select Publish.

Once published, you can share your models with others to help train, review, or use them.