Stable Diffusion shipped weights, a licence and named limits

The process wall

On 22 August 2022, Stability AI released Stable Diffusion's weights to the public rather than keeping the model behind an API. The release announcement frames this as a commitment to open-source generative AI to expand human creativity, and states the recommended checkpoint needed 6.9 gigabytes of VRAM to run, a consumer-grade requirement rather than a data-centre one. That is the process wall this record documents: for the first time, a capable text-to-image model could run on a maker's own machine, offline, without a queue or a usage meter.

What the documents show

The model card is more specific than the announcement about what the model learned from: it names laion-aesthetics v2 5+ for the final training stage, after earlier stages on subsets of LAION-2B(en) and laion-improved-aesthetics, filtered for image size, an aesthetics score above 5.0, and a watermark-probability threshold. The card states plainly that the model was trained mainly with English captions, does not achieve perfect photorealism, cannot render legible text, performs poorly on faces and people, and that some degree of memorization of training images has been observed. The repository adds that the CreativeML OpenRAIL-M licence permits commercial use but carries specific use-based restrictions, and that running the 860-million-parameter model needs a GPU with at least 10 gigabytes of VRAM, a higher figure than the announcement's own number, a discrepancy worth noting rather than smoothing over.

Authorship and limitation

Open weights changed who could inspect and modify the model, not what the model itself knows: the documented bias toward Western, English-language imagery and the documented memorization of some training images are properties of the LAION-trained checkpoint, not of any particular prompt a maker writes. The announcement's AI-based Safety Classifier is a content filter bundled with the release, adjustable by whoever deploys the model, a moderation layer rather than a guarantee about the underlying weights. The OpenRAIL-M licence is a usage licence with behavioural restrictions, not a copyright clearance for the training images or the outputs; a maker distributing a Stable-Diffusion-based service is bound by restrictions the repository says must be passed on to end users.

What survives

What a maker can check today is unusually concrete for this release: the exact training subsets named on the model card, the licence text itself, and the specific limitations Stability and CompVis chose to publish rather than leave implicit. Editorially, any claim that a Stable-Diffusion-based tool is unbiased or fully photorealistic should be checked against the model card's own limitations section before it is repeated.

• Read the model card's training-data section before describing what the model knows.
• Compare the VRAM figures in the announcement and the repository before quoting one hardware requirement.
• Check which OpenRAIL-M use restrictions apply before building a commercial service on the weights.

The release is best read as a transfer of a specific, documented artefact, checkpoints, a licence and a stated bias profile, not as the release of a fully explained system; what is undocumented, the exact image list, remains a gap the model card itself does not close.