AI: Images from Words
It’s easy to forget, but before the world had it’s OpenAI ChatGPT moment last November, there was an earlier Dall-E2 moment. And many more since then. Users around the world were captivated by the ability to type in text, and use AI to generate never before seen images. We all remember “Pope Francis wearing a long, white puffer jacket inspired by Balenciaga”.
That ‘mini-revolution’ saw products like Stable Diffusion, Midjourney and others along with Dall-E2, show the world that there was something new happening with AI and it’s ability to conjure content out of the mass of data that came before it. All done statistically, using the massive parallel computational power of GPUs and other types of chips. Companies like Midjourney in particular, continue to introduce new capabilities at a furious pace.
That revolution continues afoot in the form of ‘multi modal’ AI chat capabilities that are racing to integrate images, text, video, code, and other modalities into single user interfaces like OpenAI ChatGPT, Google Bard, Anthropic Claude2, and others.
For now, millions of users go to specialized AI image generating applications to do their daily magic. And these services all leverage the immense amount of image data that’s available on the internet, as the grist for their AI image mills.
This week saw several new announcements in this area that are worth highlighting. First up, Adobe announced their new Firefly AI image application was seeing some traction:
“Adobe today announced that its Firefly web service is now available globally with support for text prompts in 100 languages — including (Klingon and Latin). The user interface is now available in 20 languages, including in German, French, Japanese, Spanish and Brazilian Portuguese.”
In addition to expanding access, the company said that Firefly has now generated over 1 billion assets on the web and in Photoshop, which the company says makes these launches “two of the most successful beta releases in Adobe’s history.”
That coincides with image powerhouse Shutteretock announcing its partnership with OpenAI yesterday:
“Shutterstock today announced that it plans to expand its existing deal with OpenAI to provide the startup with training data for its AI models.”
“Over the next six years, OpenAI will license data from Shutterstock, including images, videos and music, as well as any associated metadata. Shutterstock, in turn, will gain “priority access” to OpenAI’s latest tech and new editing capabilities that’ll let Shutterstock customers transform images in Shutterstock’s stock content library.”
“Shutterstock says that, in addition, OpenAI will work with it to bring generative AI capabilities to mobile users through Giphy, the GIF library Shutterstock recently acquired from Meta.”
The key above is the licensing deal for the content. As one might expect, content creators are asking for compensation for their images being used in AI image generation, and lawsuits are already flying back and forth. As are industry efforts to provide licensed data that powers the AI, as seen by the companies above.
The crux of the issues continue to be debated and figured out:
“Stock content galleries like Shutterstock and generative AI startups have an uneasy — and sometimes testy — relationship. Generative AI, particularly generative AI art, poses an existential threat to stock galleries, given its ability to create highly customizable stock images on the fly.”
“Contributors to stock image galleries, meanwhile, including artists and photographers, have protested against generative AI startups for what they see as attempts to profit off their work without providing credit or compensation.”
“Early this year, Getty Images sued Stability AI, the creators of the AI art tool Stable Diffusion, for scraping its content. The company accused Stability AI of unlawfully copying and processing millions of Getty Images submissions protected by copyright to train its software.”
“In a separate suit, a trio of artists are alleging that Stability AI and Midjourney, an AI art creation platform, are violating copyright law by training on their work from the web without their permission.”
“Some experts suggest that training models using public images, even copyrighted ones, will be covered by fair use doctrine in the U.S. But it’s a matter that’s unlikely to be settled anytime soon.”
“In contrast to Getty Images, Shutterstock — perhaps unwilling to hinge profits on a lengthy court battle — has embraced generative AI, partnering with OpenAI to roll out an image creator powered by OpenAI’s DALL-E 2. (The Shutterstock-OpenAI deal dates back to 2021, but the image creator didn’t launch until late 2022.) Beyond OpenAI, Shutterstock has established licensing agreements with Nvidia, Meta, LG and others to develop generative AI models and tools across 3D models, images and text.”
A lengthy, but useful explanation of the stage ahead for AI image generation going forward this year and next. Much still needs to be figured out and worked out, especially around concerns over ‘deepfakes’, but the race to normalize AI image generation as a core infrastructure capability is certainly underway. Stay tuned.