A top supplier of stock images will compensate artists who contribute training data to its image-generation service.
What's new: Shutterstock, which launched a text-to-image generator to supplement its business in licensing images, committed to sharing revenue with contributors who permit the company to use their artwork and photographs to train its model.
How it works: The image generator is based on OpenAI’s DALL·E 2 and built in collaboration with LG AI Research.
- The developers trained the model using images (and corresponding metadata) created by artists whose work Shutterstock licenses to its customers. Contributors will be able to opt out of having their images used in future training sets.
- Shutterstock will reimburse contributors an unspecified percentage of the licensing fee for each image the model generates based on the number of their images included in the training dataset. The company offers the same deal to contributors who permit it to include their work in datasets to be licensed to third parties. Contributors will receive payment every six months.
- Users who sign up for a free account can generate up to six images per day. The company charges a fee to download and use them. Users can also upload images generated by Shutterstock’s model for licensing to other customers. The company doesn’t accept images generated by third-party image generators.
Behind the news: Rival stock-image supplier Getty banned the uploading and licensing of AI-generated art in September. Getty also recently announced its intent to sue Stability AI, developer of the Stable Diffusion image generator, claiming that the model’s training set included millions of images owned or licensed by Getty, which Stability AI used without permission.
Yes, but: Shutterstock’s revenue in 2021, the most recent year reported, was around $773 million, and image generation is likely to represent a small fraction of the revenue. Meanwhile, Image generation models like DALL·E 2 are trained on hundreds of millions of images. This suggests that individual payouts for most contributors likely will be minuscule for the foreseeable future.
Why it matters: Image generation could disrupt the business of licensing stock images. Why pay for a license when you can generate a suitable image for pennies? Shutterstock is confronting the threat proactively with a bid to own a piece of the emerging market for generated media.
We're thinking: Much of the debate over how to compensate artists for data used to train image generators has focused on what’s legal. A more important question is what’s fair. Once we hash that out, legislators can get to work updating copyright laws for a digital, AI-enabled, generative world.