Stable Diffusion is released to the public

Stability AI, working with the CompVis group at LMU Munich and Runway, released Stable Diffusion under a permissive licence with weights available for download. The model was a latent diffusion system trained on subsets of LAION-5B, a web-scraped image–text dataset, and it ran on consumer graphics cards with around 10GB of memory.

The contrast with the alternatives was the point. DALL·E 2 and Midjourney, both released earlier in 2022, were hosted services with content filters and waiting lists. Stable Diffusion could be downloaded, modified, fine-tuned on new subjects, and run offline. Within weeks there were forks, graphical front-ends, a fine-tuning technique called Dreambooth applied to personal photographs, and a large ecosystem of community-trained variants. Within months the same weights sat inside commercial products from companies with no model of their own.

The release also brought the unresolved questions forward faster than any previous system. LAION had been assembled by scraping image URLs and alt text at scale, including from stock libraries and artists’ portfolios; artists identified their own names as effective prompt keywords. Getty Images sued Stability in early 2023 in both the US and the UK, and a class action by artists followed. Separately, researchers later found child sexual abuse material in the LAION-5B index, which was withdrawn and re-released after filtering.

Stable Diffusion is the clearest single instance of the open-weights argument being settled by fact rather than debate: once capable weights are public they cannot be recalled, and everything downstream — the creativity and the harm — follows from that.