Highlights:

  • Adobe Firefly, a potent AI model family for image creation, powers the new feature.
  • Utilizing non-destructive text prompts, users of Generative Fill can extend the edges of their images and add photorealistic or surreal content.

In a statement released recently, Adobe Inc. stated that it has integrated image-generating artificial intelligence directly into Photoshop, the company’s flagship image editing tool, beginning with “Generative Fill,” a feature that allows users to edit images using a simple text prompt.

Adobe Firefly, a potent AI model family for image creation, powers the new feature. Users interact with it like every other text-to-image generator on the internet. Anyone with experience creating images with OpenAI LP’s DALLE 2, Stable Diffusion, or Midjourney will have experience entering text prompts and getting vivid images back. Users will interact with it in the same manner.

Generative Fill allows users to extend the edges of their images and add photorealistic or surreal content to them without causing any damage to the original artwork. The same feature, though a similar one has previously existed in Photoshop, could be used to eradicate objects from an image, and AI seamlessly filling in the background. To add what the user requested and fit it properly, Firefly uses what is already present in the image.

Vice President of product management and product strategy for Photoshop, Pam Clark, said, “Generative Fill automatically matches the perspective, lighting, and style of your image to make previously time-consuming tasks swift and delivers results that are truly mind-blowing and delightful.”

There are countless examples. When a user asks the AI to extend an image, it will create the remaining ocean as if the photo has continued. For example, a user could click a picture of them on a beach and extend it to the side, leaving a blank canvas. Boats with the proper perspective or a pod of breaching whales could be added to the water by selecting a surface section and typing a text prompt. A person with more imagination might ask the AI to show Atlantis emerging from the ocean.

This technique could expand a small portion of an image into a whole scene. A lone flower in the grass bearing the message: “A flower in a meadow with a blue sky and a forest with mountains.” If the canvas is big enough, the AI will create an idyllic scene with the requested mountains, flowers, and forest. Following that, the user could make choices around the freshly created image and include a cabin, more animals, or anything else that caught their attention.

The feature is built into each selection tool in Photoshop, and when it is used, it creates a new “Generative Layer,” so it won’t affect the current project. Users have the option of deleting any work they have done and continuing with their workflow or merging it at a later time into their design.

The tool was also created with high-quality assets for commercial users, so when text prompts ask for something, it will be high resolution and not based on copyrighted content. This indicates that it is secure for use in marketing materials or by brands. However, Adobe cautioned that it is still in beta testing and experimental, so business interests may want to wait to incorporate it into their workflows for now.

The popularity of these tools has also caused some skepticism among critics who worry that generative AI may give rise to deepfake concerns. Especially with the relative ease with which features like Generative Fill allows for the simple addition or removal of elements from a photo to create the appearance of realism instantly. A Twitter account that claimed to be a Bloomberg news account ignited a controversy after posting an AI-generated image of the Pentagon on fire.

Adobe limits potential by forbidding users from including political figures in their images. Images of “black smoke” in front of the White House or other notable landmarks will be avoided. Additionally, users are expressly prohibited by its AI usage guidelines from creating graphic violence, nudity, and offensive or fraudulent content, among other things.