Adobe announced Firefly 3, the latest version of its Firefly artificial intelligence image generation model, and an upgrade to Photoshop's generative fill feature.

The Firefly upgrade brings significant improvements in photorealism, quick adherence, and overall control of the final image.

Firefly 3 has been trained on billions of licensed stock images to improve overall output, including more detailed labeling of lighting, structure, and style. Initially available through the Firefly website, it will be rolled out to other Adobe products in the future, along with AI video models like Sora in Premiere Pro.

In addition to generative fill, Adobe will launch generative expand for the Firefly web app. This feature will allow users to zoom in on a canvas or change the orientation of an image, and Firefly will fill the gap.

Adobe is rapidly improving the Firefly family of transformer models, adding new features such as style references and integrating them with existing Adobe products.

Firefly 3 is being learned with licensed Adobe Stock images (including Midjourney's work) and will likely be the first version with video capabilities later this year.

While most people will interact with Firefly through generative fills in Photoshop and template and other content creation in Illustrator and InDesign, Adobe is investing in a Firefly web app.

Text-to-image models can also use existing images and copy their style and structure. For example, if there is a perfectly lit photo, its style can be copied, or if there is a product image, its structure can be copied but its content can be changed.

"In just over a year, Firefly has become the image generation tool that millions of creators use every day to develop their ideas.

"We continue to advance the state-of-the-art with the Image 3 Foundation Model and can't wait to see how our creative community pushes the boundaries of what is possible with this beta build.

I tried it out with a handful of prompts and compared it to Firefly 2, which was already impressive in art and design-based creation, etc. Firefly 3 is a huge step forward in photorealistic imagery. I think it will be a blow to the stock image field.

According to Adobe, working with Firefly 3 is focused on speeding up the ideation process, allowing designers to go from idea to full-blown image in as little time and with as little friction as possible.

I think they accomplished that. Unlike Midjourney, where you have to learn multiple parameters and how to implement them, Firefly has a clearly defined set of menu options.

Firefly 3 seems to have improved photorealism, offering a wide range of output from a single prompt across photography, art, and illustration styles, with options to set mood and lighting.

According to the company, the understanding of the prompts has also improved, and when we asked it to display an old castle by a lake and people boating, it hit the mark beautifully. In other words, relatively vague prompts should be avoided and details should be accurate.