• The ability to generate an image by typing several textual instructions brings home the potential of AI and seems indistinguishable from magic.
• Stability AI was released in August 2022. Stability AI differentiates itself from competitors by not censoring the user’s creation.
Stability AI (SAI) was co-founded in 2020 by Emad Mostaque, who earned a master’s degree at Oxford University in mathematics and computer science. After a stint as an analyst at several different hedge funds, Mostaque decided to switch gears and pursue his interest in AI. Mostaque provided the initial funding but soon found the company awash in venture funding, giving SAI a valuation of approximately $1 billion. With SAI, artificial intelligence moves beyond analytics and becomes a creator. In order to create an image with SAI, a creator simply types descriptive text into the DreamStudio, and an image slowly appears.
In October 2022, SAI held a press conference introducing itself to Silicon Valley, California (US). However, even with a press conference understanding the lineage of SAI is not easy. Stability AI is the brains behind Stable Diffusion, which is a free and open-source text-to-image generator launched in August 2022. Stable Diffusion software was created by German academic researchers and Patrick Esser, a researcher who currently works with Runway, an SAI competitor.
Not surprisingly, Runway also uses Stable Diffusion to enable some of its current and upcoming features, including AI-generated video. Runway released a new version of Stable Diffusion on Hugging Face, a site that hosts machine-learning models. In response, SAI sent Hugging Face a notice requesting the company take down the model, claiming Runway’s action was a leak of their intellectual property (IP). In response, Runway wrote that the AI model was released under an open license and there had been no breach of IP. Stability’s legal team later withdrew the takedown request. Hoping to avoid a stain on its corporate image, SAI approached Esser and offered to donate computing power to improve Stable Diffusion further. Esser supervised the new training of the model, funded by SAI, which improved the software’s understanding of the text as well as the quality of the images it could produce. However, the incident has left many developers bitter over the entire experience.
In order to run SAI’s text-to-image application, the company has a cluster of more than 4,000 Nvidia A100 GPUs running in AWS, which it uses to train AI systems, including Stable Diffusion. However, training has opened SAI to a litigation tsunami, as it did not filter from its training data copyrighted work.
SAI’s refusal to impose filters to weed the creation of offensive images as well as ‘deep fakes’ has already caught the attention of the US Congress. However, while SAI has released its text-to-image software to the general public free of charge, SAI looks to monetize its capabilities by selling the software and add-ons to businesses. SAI may have to address the filtering issue to generate revenue, as no company wants to use software that could cause harm.