Adobe has unveiled a new feature inside its Adobe Firefly portal that gives users more control over the final look of images and art created using Generative AI. The feature is called Structure Reference and offers users control over an aspect of their output that has been unavailable previously.
Firefly allows users to create images and art by typing in a text prompt that describes the desired result. To illustrate the process, I input the following prompt: "Portrait of a dark-skinned woman with shoulder-length dreadlocks on a beach at sunset.” In a checkbox option, I selected Photo, rather than Art, for the final result. The program quickly created several images that matched the description.
Before exploring the addition announced today, let's have a look at the Style Reference feature that has been available previously. To help guide the program in the direction of how I wanted the final image to look, I uploaded an image of a model with dreadlocks named Lyric that I had photographed a few years back.
Firefly referenced this photograph to create new images using the same prompt I had initially entered. As you can see, the model in the latest images has a similar skin texture and color palette as the photograph of Lyric. The new feature introduced today, Structure Reference, works similarly to the Style Reference feature.
I spoke with Alexandru Costin, Vice President, Generative AI at Adobe, about this addition. He explained that a user can upload a reference image that the program will utilize in creating the design and structure of a photograph or artwork. Since Adobe’s Generative AI is designed for personal and commercial use, Alexandru quickly pointed out that Adobe’s AI has been trained on data owned by Adobe. “We do it responsibly, not only training on data that we own, which guarantees the generation of safe for commercial use content, but also doing it transparently with minimized biases, minimized potential harm, and in full transparency with our internal policies and externally driven content authenticity,” said Alexandru.
In a demo performed for me, Alexandru input the prompt, “A stack of pancakes with maple syrup, berries, and a cube of butter,” with the output set to Art (not Photo). These were some of the results.
Then, Alexandru uploaded a black-and-white sketch of a stack of books to be used a Structure Reference.
As you can see from the final images below, the pancakes now have a structure suggested by the reference image.
For commercial users, Structure Reference, when used alongside Style Reference, can help maintain a consistent look for artwork or images used in marketing. “If you have a particular Fall campaign, and if you want to generate new content in that style, you can fine-tune and create a custom Firefly model that generates content in that style so you can guarantee that you don't need a lengthy review process and content regeneration. We call that the Firefly Services. It is a collection of APIs that enterprises can use to generate new images or expand existing images, to mix images with Photoshop templates, to maintain brand consistency.”
Structure Reference gives the user the ability to modify an AI-generated drawing or photo by influencing the structure of that creation. Or, it can be used to transform a simple sketch into something more complex and detailed by adding elements such as depth, texture, lighting, and color. As with any new technology, a user must be willing to experiment and even play around with the interface to create the desired results. This ability to use text to create art or photographs has only been around for about a year, but the creations are impressive. The future of this tech is promising for not just major corporations but amateurs and professionals as well.
