Adobe has new generative AI features for rookie Photoshop users, but there are concerns from creative professionals.
On Tuesday, Adobe announced a beta version of Photoshop with new features to generate and edit images from text prompts. Generative AI capabilities are powered by Adobe’s Firefly Image 3 model which was also released today.
New to the photo-editing software is a Generate Image tool that creates an image from a text prompt, giving people feeling daunted by a blank page a jumping off point.
Credit: Adobe
Generative Fill, a preexisting tool that fills out the background or expands an image with a broader scene, now includes the ability to build off of an image with a Reference Image feature. So if you want to create an image with a specific aesthetic you can upload an image and nudge the tool to work off that aesthetic or motif.
Credit: Adobe
For Photoshop novices (and pros for that matter) these tools make it incredibly easy to bring their ideas to life. Especially now that Firefly Image 3 is capable of amazing photorealistic detail and improved understanding of prompts. But Adobe has pushed ahead with generative AI features despite controversy around its training data and backlash from creative professionals.
The issue of training data for AI models is a murky one that isn’t going away any time soon. Companies like Google and OpenAI have been sued for training their models on copyrighted works, and that’s just for text outputs. The unveiling of OpenAI’s Sora (which isn’t available to the public yet) and AI video generators like the one Adobe just released have expanded the issue to other mediums. While companies are charging forward with AI advancements to meet their competitors, many creatives like @Rahll on X, are outraged by the lack of oversight over technologies that threaten to replace or overhaul their vocations.
Adobe has been adamant about Firefly being safe for commercial use, and has touted its model as the ethical alternative to competitors like Midjourney or Stable Diffusion that have been accused of training their models on copyrighted works without credit and compensation. Yet a recent Bloomberg report found that Firefly was trained on AI-generated images derived from Midjourney and other rivals, meaning Adobe’s purported “commercially safe” model contains training data drawn from works perceived by many as having been infringed upon or stolen.
A spokesperson said Adobe’s generative AI models are trained on “licensed content, such as Adobe Stock, along with public domain content.” But, according to Midjourney enthusiast Nick St. Pierre, around 13 percent of Adobe’s stock library is AI-generated content which is part of the data used to train Firefly.
That said, the spokesperson said “the training data must comply with our IP guidelines and pass our multi-layered, continuous review and moderation process,” so there is ongoing moderation of unlicensed AI-generated content that might slip through the cracks.
Particularly when it comes to the Reference Image tool, X users have pointed out the issue of creating an image based on a photographer or artist’s work. “Awesome. Steal ‘reference images’ from real photographers who went to great lengths and great expense to make them,” posted @notbrodyjenner. “Then pick an artist who spent their career creating a unique style, rip it off wholesale and make whatever ai garbage you want. Awesome.”
The spokesperson said Adobe compensates contributors through Stock royalties and through a bonus program that incentivizes users to contribute to Firefly training.