Andrew Heinzman
It had to happen eventually. Adobe is opening a public beta for its new “family” of AI image-generation tools, called Firefly. Many of these tools rely on a text input, similar to DALL-E. That said, Firefly is notable for several reasons, including the way it handles copyright.
Adobe is launching two browser-based Firefly tools today. First, there’s the Text to Image tool, which spits out an image based on your prompt (and allows you to modify the image using a suite of “Style” and “Concept” controls). And then there’s Text Effects, a tool that creates stylized text and fonts based on natural language processing—Text Effects lets you ask for some pretty wacky stuff, including fonts that look like melting chocolate.
Other Firefly tools will arrive in the future. And, naturally, Adobe plans to integrate this stuff with its desktop apps. Some of the Firefly tools teased by Adobe include an image upscaler, an automated paintbrush creator for Photoshop, a more advanced raster-to-vector algorithm in Illustrator, and a variety of image-editing tools that build on things like content-aware fill.
As for Firefly’s training data, Adobe claims that it will respect artists’ copyright. All images utilized by Firefly are licensed for training, available in the Adobe Stock library, or public domain. Adobe is also searching for a way to compensate Adobe Stock unloaders for training data, similar to what Shutterstock is doing with its AI program.
Additionally, Adobe wants to create a “not for training” tag that can be applied to customers images (specifically images made without AI). This could reduce copyright concerns across the generative AI industry, assuming that other companies and individuals actually respect the “not for training” tag. Presumably, such a tag would be included in an image’s metadata (which may deter automated crawlers from grabbing images). Adobe may develop a similar tagging system to help identify AI-generated images.
While I appreciate Adobe’s approach to copyright, I do have one concern. Adobe wants to give customers the ability to build a dataset using their own content. This could allow artists or businesses to create images that fit their existing style or brand. But what’s stopping me from taking someone else’s art and dumping it into a custom dataset? (For what it’s worth, this feature is a long way out, and Adobe says it’s trying to find a way around this problem.)
You can request access to the Firefly beta at Adobe’s website. Note that Adobe will host a Firefly conference at 4PM EST on March 21st. We will update this post with any new information revealed during the conference.
Source: Adobe
