In late August 2022, Stability AI released Stable Diffusion, an image generation model based on diffusion techniques, as open source. The model produced images of comparable quality to OpenAI’s DALL-E 2, which had launched a few months earlier and had been available only through a controlled API. The difference was that Stable Diffusion ran locally. Anyone with a reasonably capable GPU could generate images on their own hardware, with no API limits and no upstream provider deciding what they could and could not generate.
The technical achievement was significant on its own. Diffusion models had been improving rapidly through 2021 and 2022. Stable Diffusion put a state-of-the-art version in the hands of anyone who wanted to download it. Within days of the release, ports to consumer-grade hardware appeared. Within weeks, Stable Diffusion was running on Macs with M1 chips, on gaming PCs, and in cloud notebooks accessible to anyone.
The cultural reaction was layered. Artists who had been skeptical of AI image generation found themselves contending with a much wider deployment of the capability than they had expected. The training data for Stable Diffusion included a large amount of copyrighted artwork that had been scraped from the open web. The legal and ethical questions about that training data had not been resolved before the release. The release made resolving them retroactively much harder.
The application ecosystem developed quickly. Within months, hundreds of fine-tuned variants of Stable Diffusion existed, many specialising in particular styles or subjects. Tools for editing existing images using the model emerged. Plugin-style integrations into existing creative software appeared. A category of work that had not been thought of as automatable was suddenly subject to a tool that could do significant parts of it in seconds.
The wider implications, for stock photography, for illustration as a profession, for online media, for the legal status of training data, were raised but not resolved by the release. They would continue to play out for years.
What Stable Diffusion settled was the question of whether high-quality generative image AI would be a controlled capability available through major providers or a widely available capability that anyone could run. The release made the second answer unavoidable. Everything that followed had to assume that anyone with a moderately capable computer could generate any image they wanted, for any purpose, without anyone else being able to prevent it.