Best practices for creating custom models (beta)

Learn about the best practices to effectively prepare and create custom models.

Pick a strong use case

Use quality images to train your model

• Use JPG or PNG files.
• Choose at least 10-30 high-quality images that showcase the brand-specific styles and concept subjects you want to achieve.
• Capture a varied set of images representing the style or subject.
• Ensure that each image file size does not exceed 50 MB.
• Ensure the images have a resolution higher than 1024x1024 pixels with a maximum 16:9 aspect ratio for landscape or 9:16 for portrait. 
• Keep the aspect ratio consistent with the training dataset. If the training set is in portrait, and you generate square images, they will have cut-off issues upon generation. 
• Crop your sample images to focus on the most important visual elements. For example, exclude images that show a person or character off in the distance with a small face or body.
• Include images displaying various viewpoints and backgrounds while maintaining a consistent aesthetic.
• Make sure your images don’t include an unintentional pattern that you do not want, such as having a white background in every image.
• Remove distracting elements that you do not want the model to learn, such as a collage in the background of a portrait or a hat on a character.

Review Model Tags

• Include 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.
• Include a minimum of three Model Tags.

Review Captions

• Use captions to enhance detail and train the custom models on concepts you want the model to generate. 
• Keep image captions concrete and descriptive, using language that you will use when prompting with the model.
• Vary sentence structure across all your image captions.
• Modify auto captions as needed to inform the model of the details of the concept.
• The Firefly base model does not know famous people or places, so captions should include descriptions of these places to improve potential outcomes.

Use crisp prompts that align with your training data

• Include similar words and phrases in your prompts that you used in captions.
• When writing your prompt, use concepts that relate closely to the data used to train the model. Prompting with unrelated or new concepts can lead to unsatisfactory image output. For example, asking for a black-and-white illustrated rocketship from a model trained on colorful lifestyle photography.

Use advanced generation settings to further refine your image

You can control advanced generation settings in the left panel of Firefly and by selecting Generation settings in the prompt bar of Boards.

For custom models: 

• The Visual Intensity slider is set to the lowest by default for optimal identity preservation. However, for creative use cases such as Style reference, increasing visual intensity can produce more vibrant results.  
• When using  Composition references for subjects, opt for images with white backgrounds or sketches depicting the subject in the desired pose.