But these tools have raised serious legal questions about the copyright protection for the assets that these AI models are trained on.
“Every piece of content that we train on is something that we have acquired the license of, or that is published under a verifiable and known license,” says Greenfield.
This approach does come with some limitations. If Firefly were asked to generate an image of a Disney cartoon character, like say Mickey Mouse, “it would do a horrible job of it,” concedes Greenfield. “And that’s by design and on purpose.”
Greenfield says that AI tools based on every image found on the internet produce less desired outputs, not just for potentially infringing on IP, but because it is representative of a vast trove of data that doesn’t always have the best quality. “There’s the raw science of how you build the model, but a massive amount of work goes into data curation and preparation,” says Greenfield. “The average piece of content on the internet isn’t necessarily what you want to put in your ad.”
Since Firefly’s launch, Adobe has had to make some modifications to the images in the company’s asset bank. Early on, generative AI wasn’t great at producing clear images of hands, so Adobe had to reach out to the photographers it works with to get more licensed pictures of hands to train the AI properly.
All Firefly content also goes through a moderation process that includes a mix of human and computer oversight, eliminating harmful images, but also those that may contain sensitive IP. A photographer may have exclusive license to an image that they produced, but if there’s a trademark asset like a Nike Swoosh or Starbucks Siren logo, Adobe will nix the image.
More recently, Adobe has integrated image and video models from OpenAI, Google, Pika Luma AI, and Runway into the company’s Firefly app.
For Adobe, Greenfield says pulling in these partnership models reflected an evolution to how creative professionals are working with AI today. He says that customers want access to a wide variety of AI models, especially as these technologies quickly advance. This is similar to the multi-modal approach most CTOs and chief information officers have embraced when deploying AI coding tools for software developers or the application of other uses of AI in marketing, legal, and communications to improve worker productivity.
Adobe has added content credentials to make it clear to marketers when the assets they are creating are safe to use for commercial production (with Firefly) versus for ideation purposes (the external partner models). Customers have the final say on what path works best for them.
“We have a lot of customers who have different opinions on when to use different types of models and how they feel about commercial safety,” says Greenfield. “A lot of them feel that in ideation, they’re open to using anything.”
John Kell
