The Stable Diffusion Release

The event was not the model. Text-to-image generation already existed in 2022, behind interfaces. The event was a file: on 22 August 2022 Stability AI published the trained weights of Stable Diffusion for anyone to download — "It is our pleasure to announce the public release of stable diffusion following our release for researchers," twelve days after the 10 August announcement had described the system and promised the release was coming.

Simon Willison, writing a week later, put the distinction in one sentence: "It's similar to models like Open AI's DALL-E, but with one crucial difference: they released the whole thing." You could, he noted, download and run the model on your own computer and use it for commercial and non-commercial purposes. The license was not classical open source but a new instrument, CreativeML OpenRAIL-M — weights free to use and redistribute, with use-based restrictions written into the license itself. The package shipped with "an AI-based Safety Classifier included by default," and the announcement was upfront about what a local model means: the classifier is part of the software, not a wall around a server.

The thing being handed out was small. Stable Diffusion is "a latent text-to-image diffusion model" — the architecture from "High-Resolution Image Synthesis with Latent Diffusion Models," posted by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Björn Ommer of the CompVis group on 20 December 2021 and presented at CVPR 2022. It pairs an 860M-parameter UNet with a 123M-parameter frozen CLIP text encoder, and was trained on 512x512 images from subsets of the LAION-5B dataset. Under a billion parameters in the UNet: the image generator that changed the internet's visual culture is smaller than GPT-2. At release, Stability put the memory footprint at 6.9 GB of VRAM — which is the load-bearing number, because it meant a consumer gaming GPU was enough. "Stable Diffusion runs on under 10 GB of VRAM on consumer GPUs, generating images at 512x512 pixels in a few seconds," as the August announcement put it.

What open weights buy is visible in the first week's record, which Simon Willison compiled on 29 August: "In just a few days, there has been an explosion of innovation around it." Web interfaces, a Photoshop plugin demo, video-frame conversion, and above all img2img — feeding the model an image plus a prompt instead of a prompt alone — turning a text toy into an iterative illustration workflow, built by people who did not need permission because there was no one to ask. The weights lived on Hugging Face under the CompVis organization; the fine-tunes, forks and interfaces that followed are the reason a "checkpoint" is now something hobbyists trade.

The other half of the record is the argument the release started. Emad Mostaque's launch line — "We are delighted to release the first in a series of benchmark open source Stable Diffusion models that will enable billions to be more creative, happy and communicative" — is a claim about who should hold generative capability, and every objection to it arrived on schedule: trained-on-scraped-art grievances, safety-filter removals, and the question of what a license clause is worth once a file is on a million disks. Those disputes have their own records. What belongs to this event is the demonstration itself: the day image generation became something you have, rather than something you visit, and the proof of how much engineering a community will do in a week when the artifact is simply given to them.