Edit in GitHubLog an issue

Firefly Image 3 Model

The Firefly Image 3 Model offers an improved generative Al image model for faster ideation and content creation.

Overview

This document provides an overview of the changes and new APIs introduced for the Firefly Image 3 Model.

Specific benefits include:

  • Improved generation quality with better depiction of details and more photorealistic results.
  • Greater variety in generated outputs from a single prompt for photos, art, or flat illustrations.
  • Improved ability to take reference images into account, both for stylization and for structure matching.
  • Better understanding of prompts and compositions of complex scenes and text rendering.
  • Improved image variations with varied options for composition, mood, and lighting.

See this link for more details on what's new.

New APIs

The following new APIs were introduced:

  • Generate Similar Images: Pass in a source image to use as a reference for generating similar image results.
  • Generate Object Composite: Upload an image (with or without mask), such as a product photo, and utilize a text prompt to generate a seamlessly composited scene featuring the product. Additionally, provide optional parameters to influence the style and content of the background. See this link for more specific details.

Updated APIs

The following existing APIs were updated to use the new Image Model 3, and have new and/or renamed parameters:

  • Generate Images API:

    • The n parameter has been renamed to numVariations.

    • The styles object parameter was renamed to style and contains renamed child parameters as well. You can still supply a reference image to be used, but it should now be passed as url (for a presigned image URL), or uploadId for an uploaded image within the imageReference.source field:

      Copied to your clipboard
      style": {
      "presets": [
      "bw"
      ],
      "imageReference": {
      "source": {
      "url": "https://example.com/",
      "uploadId": "string"
      }
      }
      "strength": 50,
      }
    • The photoSettings parameter has been removed.

    • The locale parameter has been renamed to promptBiasingLocaleCode.

  • Expand Image:

    • The n parameter has been renamed to numVariations.

    • locale has been renamed to promptBiasingLocaleCode.

    • The image parameter now has a source and a mask that can be provided as sub-parameters in the form of either an uploadId of an uploaded image (ie: image.source.uploadId or image.mask.uploadId, or a url for a presigned URL (ie: image.source.url or image.mask.url):

      Copied to your clipboard
      "image": {
      "source": {
      "url": "https://example.com/",
      "uploadId": "string"
      },
      "mask": {
      "url": "https://example.com/",
      "uploadId": "string"
      }
      }
    • A new placement object parameter can be specified, with insets and alignment child parameter values that can be provided. See the placement section below for more details.

  • Fill Image:

    • The n parameter has been renamed to numVariations.

    • locale has been renamed to promptBiasingLocaleCode. ​- Support for a negativePrompt parameter has been added.

    • The image parameter now contains both the source parameter for the image input, and the mask parameter for the mask to use. (The previous version of the API had mask at the same level as the image parameter). Please note, for both, you can supply either a url (for a presigned URL – ie: image.source.url or image.mask.url) or an uploadId (for an uploaded image – ie: image.source.uploadId or image.mask.uploadId):

      Copied to your clipboard
      "image": {
      "source": {
      "url": "https://example.com/",
      "uploadId": "string"
      },
      "mask": {
      "url": "https://example.com/",
      "uploadId": "string"
      }
      }

    See this link for more details about the capabilities available for expanding and filling images with Firefly Image Model 3.

Next steps

  • Privacy
  • Terms of Use
  • Do not sell or share my personal information
  • AdChoices
Copyright © 2024 Adobe. All rights reserved.