one year on
Stability AI releases Stable Diffusion to the public with local GPU support
The open-weights release of Stable Diffusion lets anyone run a powerful text-to-image model locally with a permissive license, a default safety classifier, and DreamStudio for commercial access.
On August 22, Stability AI releases Stable Diffusion, an open-weights text-to-image model that generates 512×512 images in seconds on a consumer GPU with 6.9 GB of VRAM. Released under the Creative ML OpenRAIL-M license, the model allows commercial and non-commercial use, and includes a safety classifier to filter undesirable outputs. The release is the result of a collaboration with HuggingFace and CoreWeave.
The model’s open nature marks a shift from the API-gated approach of DALL-E 2 and MidJourney. Since the weights are freely downloadable, users can run the model locally, modify it, and build derivative projects. Stability AI also launches DreamStudio, a commercial web service for generating images, and promises future optimizations for AMD and Apple M1/M2 chips.
The model is released under the Creative ML OpenRAIL-M license, a permissive license that allows commercial and non-commercial usage and focuses on ethical and legal use of the model. It includes by default an AI-based Safety Classifier that understands concepts and other factors in generations to remove undesired outputs. Stability AI requests that users read the model card for limitations and use the system in an ethical, moral, and legal manner.
The record
One year later — open only if you can handle spoilers
Stable Diffusion's open release catalyzed a wave of innovation, spawning countless fine-tuned models, tools like Automatic1111's web UI, and foundational models like SDXL. It also ignited the open-versus-closed AI debate that continues to define the industry, and its licensing model influenced later releases such as Meta's Llama.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy