Firefly Training Included Rivals’ AI Images

Firefly Training Included Rivals’ AI Images

Adobe made users believe it used entirely its own image stock to train its AI image generator Firefly model, but reports show it also included AI images from rivals like Midjourney.

Firefly is touted as the most ethical and commercially safe image generator using its own image stock. But it has been discovered they also use AI generated images by tools like Midjourney contrary to claims of using Adobe Stock.

Not so ethical AI

The company has sold its Firefly as some kind of an “ethical AI” which trains on Adobe’s own library of licensed stock images. These are millions of designs, images, illustrations from the Adobe Stock library.

Contrary to this, a new report has emerged showing the image generator is not so clean after all, as the image it projects on the market. A Bloomberg report has revealed the Photoshop maker has “a few skeletons in the closet.”

Some of the images used during training of Firefly were sourced from competitor Midjourney, which has not revealed the source of its training data although many suspect the images were obtained through scrapping the internet without any licensing. This is despite Firefly being seen as “commercially safe.”

For Rachel Metz, this is not about “AI incest” as it is now known, but a contradiction of what Adobe projects its AI image generator to be in the eyes of the public.

“It’s about making a deliberate choice to include AI generated images from Adobe Stock in its dataset while publicly positioning itself as very different from the companies whose tools were used to make those AI-generated images,” Metz responded on X.

Some users however feel avoiding data contamination is close to impossible.

“Literally impossible to not have any training data contamination unless you manually attribute and verify each image which is way too cumbersome,” responded Raj Singh.

Also read: Gen AI Music Hits 20% Success Rate in Commercial Briefs

The AI-labelled images

The Bloomberg report further reveals that Adobe Stock, which was used to train Firefly has 57 million images labelled AI-generated, which is about 14% of the total images in that database.

The Photoshop maker has revealed that only 5% of the images that were used in training Firefly falls into this category while the rest “were part of Adobe Stock Library, which meant they’d been through a ‘rigorous moderation process’.”

“Every image submitted to Adobe Stock, including a very small subset of images generated with AI, goes through a rigorous moderation process to ensure it does not include IP, trademarks, recognizable characters, or reference artists’ names,” a spokesperson told Bloomberg.

At its launch, Adobe said Firefly was against copyright theft, offering an indemnity to its commercial customers as the company convinced them its AI model is safe.

Better than its competitors

Adobe also projected its AI image creation tool as a “safe alternative” to competitors Midjourney and Dall-E because, it claimed the data was licensed and “cleared for use in training the model.”

According to Tom’s Guide, not all artists were that keen at the time and “felt they were coerced into agreeing to let their work be used by the creative tech giant.” However, the idea sold was that images made with Firefly were safe to use “without risk of being sued for copyright theft.”

The Bloomberg report however has raised another dimension to Adobe Firefly’s “ethicality.”

Adobe has differentiated its AI image generator from peers like Dall-E 3, Stable Diffusion, and Midjourney declaring “commercially safe.”

“Adobe has established the AI ethics principles of accountability, responsibility, and transparency,” wrote the company in one post.

The company went on to create a compensation plan for all artists whose work went through into the first training of Firefly. Some of the artists, according to According to Entrepreneur Daily, submitted images that were created by Midjourney but were compensated by Adobe for their input.



Image credits: Shutterstock, CC images, Midjourney, Unsplash.