Getty launches 'safe' text-to-image AI tool

Getty launches ‘safe’ text-to-image AI tool

Source Node: 2293243

Getty Images announced its own text-to-image generative AI tool on Monday, insisting it is “commercially safe” as it’s been trained exclusively on Getty’s own stock photo platform.

Dubbed “Generative AI by Getty Images” the service uses Nvidia’s Edify model to generate images, videos or 3D graphics. Edify is part of Nvidia’s Picasso platform, which hosts a range of generative AI models on its cloud service.

Unlike other commercial text-to-image tools, Getty’s model is not trained on third-party images scraped from the internet. It was only fed data that the stock photo provider owns – meaning any pictures generated by the software shouldn’t violate any copyright laws.

“We’ve listened to customers about the swift growth of generative AI – and have heard both excitement and hesitation – and tried to be intentional around how we developed our own tool,” said Grant Farhall, chief product officer at Getty.

“We’ve created a service that allows brands and marketers to safely embrace AI and stretch their creative possibilities, while compensating creators for inclusion of their visuals in the underlying training sets.”

The service’s output can be used under Getty Images’ standard license. Users will pay a one-off fee to Getty for licensing the AI-generated image, but cannot license it to third parties or receive royalty payments on machine-made works.

“Similar to when customers license content from Getty Images’ pre-shot library, they do not own the file [or] image itself, but Getty Images grants customers a license to use the content,” a Getty spokesperson told The Register. “As such, customers are not able to further license generated content to additional parties, nor receive their own royalties as part of their agreement.”

Fees will vary depending on the number of times a customer runs the model to generate their desired image. Given an input description, or prompt, they will see four different variations of images generated by the software.

“As customers commit to higher volumes of prompts, they will achieve savings on a per prompt basis. Customers will also be able to download a licensed version of any visual they generate, for no extra charge,” Getty’s spokesperson told us.

The image library said it will not add people’s custom AI-generated images to its own database, but will retain the content as well as the input descriptions, or prompts, used to create it, to train its model.

“Your prompts, generated images, and feedback will all be used in retraining our model to become more effective for you, our customer. By allowing our model to learn and improve through retraining, we can provide you with an improved experience such as better understanding the intent of your prompt or reducing deformations in generated content,” the image org confirmed in an FAQ.

A content filter analyzing input prompts will block users from attempting to generate inappropriate or lewd content, the picture biz said. Getty believes that the model should be safer to use in such situations than other commercial tools on the market, in part thanks to its previous curation efforts for the library used to train its AI.

Last year, Getty banned people from uploading AI-generated artwork onto its platforms because it didn’t want to host images that could potentially infringe intellectual property rights. The biz has also sued Stability AI, a London-based startup, for scraping its images without consent. The scraped images were used to train Stability’s own text-to-image tools.

Getty is not the first to make a pitch at the legally safe AI market. Earlier this month Microsoft announced it would cover any copyright cases filed over the generation of software by CoPilot – for paying customers, at least. ®

Updated to add

Getty has clarified its position on paying creators in a statement to The Register.

“We will compensate contributors for all licensing. This includes where contributors’ content is used as training data for AI,” a spokesperson told us.

“On an annual recurring basis, we will share in the revenues generated from the tool with contributors whose content was used to train the AI Generator, allocating both a pro rata share in respect of every file and allocating a share based on traditional licensing revenue.”

Time Stamp:

More from The Register