When the model runs on a laptop
Stability AI released Stable Diffusion this week under a permissive license, small enough to run on a consumer graphics card. Once anyone can run the model, what’s designed around it is the product.
On August 22, Stability AI released Stable Diffusion to the public. The license, CreativeML OpenRAIL-M, allows commercial and non-commercial use. The model runs on a consumer graphics card with under 6.9 GB of video memory. It ships with an AI-based safety classifier switched on by default, with settings people can adjust.
In April, the most impressive image model anyone had seen was invite-only. This week, a comparable one is a download.
The model as a material
Once anyone can run the model on their own machine, the difference between products is what’s designed around it.
That changes the question product teams ask. When a model is scarce, the product is access to it, and the interface can be a text box and a button. When the model is a file anyone can download, the text box is worth nothing on its own. What’s left to compete on is what the product does with the model: the editing tools, the workflow it fits into, how it handles the result that’s almost right, and whether it remembers what you were trying to make.
I’ve seen a version of this before. The statistics behind an experimentation platform aren’t secret. Anyone can compute whether a change moved a metric. What made Uber’s platform useful was what we designed around the math: one page per launch, a sentence saying what’s wrong, and a rollback you can aim at the segment that’s hurting. The math was necessary. The design was what people used.
A default you can turn off
The release also shows what happens to safety when the model leaves the server. Stability’s classifier is on by default, which is the right default. But on your own machine, a default is a suggestion. Anyone who wants it off can turn it off.
That moves responsibility in two directions. Products built on the model now carry the guardrails themselves, because the platform can’t enforce them. And the questions about what the model was trained on, and whose work it learned from, follow every copy, as well as the company that trained it.
I don’t think the release can be undone, and I’m not sure it should be. It does mean that what used to be a policy decision on one company’s server is now a design decision in every app that ships this model.