Adobe’s Firefly, a family of generative AI models, has received criticism from creatives in the past. The Firefly image generation model, in particular, was considered underwhelming compared to its competitors. However, Adobe is addressing these concerns with the release of its third-generation model, Firefly Image 3. This new model, available in Photoshop (beta) and the Firefly web app, promises more realistic imagery and improved capabilities in understanding complex prompts and scenes.
According to Adobe, Firefly Image 3 can accurately render typography, iconography, raster images, and line art. It also excels at depicting dense crowds, people with detailed features, and a variety of moods and expressions. In a brief unscientific test, Image 3 appeared to be a step up from its predecessor, Image 2.
One noticeable improvement in Image 3 is its lighting capabilities. A comparison of headshots from Image 3 and Image 2 reveals that Image 3 produces more detailed and lifelike results with better shadowing and contrast. Similarly, when comparing scene understanding, Image 3 outperforms Image 2 in terms of detail and expressiveness. While there may be some minor issues with certain elements in Image 3, overall, it showcases significant improvements.
These advancements in Image 3 can be attributed to a larger and more diverse training data set. Like its predecessors, Image 3 is trained on uploads to Adobe Stock, as well as licensed and public domain content. To address ethical concerns, Adobe pays contributors to include their content in the training data set. Additionally, Adobe trains Firefly models on AI-generated images, which some consider a form of data laundering. However, Adobe defends this practice, stating that AI-generated images make up only a small portion of the training data and go through a moderation process to ensure they comply with copyright regulations.
While diverse training data and content filters aim to provide a better user experience, they do not guarantee a flawless outcome. Users have previously generated inappropriate content with Image 2, highlighting the need for ongoing community feedback and improvement.
Firefly Image 3 introduces several new AI-powered features in Photoshop. The “style engine” and auto-stylization toggle allow for a wider array of colors, backgrounds, and subject poses. Reference Image enables users to condition the model on an image to align the generated content with desired colors or tone. Additionally, three new generative tools—Generate Background, Generate Similar, and Enhance Detail—leverage Image 3 to perform precision edits on images. These features have been in beta in the Firefly web app and are now making their debut in Photoshop.
Adobe is also focusing on improving its Firefly web app. With the release of Image 3, the web app now includes Structure Reference and Style Reference. Structure Reference enables users to generate images that match the structure of a reference image, while Style Reference allows for style transfer while preserving the content of an image.
Despite these upgrades, Adobe has stated that pricing for Firefly image generation will remain the same for now. They will continue to offer various tiers and a generative credit system. Adobe’s indemnity policy, which covers copyright claims related to works generated in Firefly, will also remain unchanged. The company will continue watermarking AI-generated content and attaching Content Credentials metadata.
In conclusion, Adobe’s Firefly Image 3 represents a significant improvement over its predecessors. With enhanced capabilities, a larger training data set, and new AI-powered features, Adobe aims to provide creatives with more realistic and versatile generative AI models. While challenges and ethical concerns persist, ongoing feedback and community involvement will play a crucial role in refining these technologies further.