Overview
Stability AI develops the Stable Diffusion family of image generation models along with related generative models for video and audio. Its models are released as open-weight models: the trained weights are published for anyone to download and run, most commonly through Stability AI's Hugging Face organization, with inference code published on GitHub.
Model and technology information
Stability AI's released model lines include multiple generations of Stable Diffusion for text-to-image generation, up through the Stable Diffusion 3.5 release (offered in Large, Large Turbo, and Medium variants at different parameter counts and hardware requirements) as well as Stable Video Diffusion for image-to-video generation, and Stable Audio for text-to-audio generation. The image models are built on latent diffusion architecture, generating in a compressed latent space rather than operating on raw pixels directly.
Open-weight licensing approach
Stability AI distributes its core models under a Community License: free to use for research, non-commercial purposes, and for commercial users generating under a stated annual revenue threshold of USD $1 million. Organizations above that threshold require a separate, paid Enterprise License, which adds implementation support and options like custom model training. Under the Community License, outputs generated from the models belong to the user, and community fine-tunes (further checkpoints or LoRAs built on top of the released weights) are permitted.
Content policy and acceptable use
Stability AI's published Acceptable Use Policy prohibits a defined set of uses across its own products and services: generating child sexual abuse material or other content sexualizing minors, non-consensual intimate imagery, "illegal pornographic content," and content depicting sexual intercourse or sexual violence, alongside extreme gore, weapons-related content, and political-campaigning uses. Stability AI states it reports child sexual abuse material to the National Center for Missing and Exploited Children (NCMEC).
This acceptable-use policy governs Stability AI's own products and services. Because Stability AI's base models are released as open weights, they can also be downloaded and run independently through other software and platforms, which operate under their own separate policies rather than Stability AI's.
Safety and content authentication
Stability AI describes a "safety by design," open-first approach to its models. Its published safety principles cover applying filters to training data to remove unsafe images before a model is even trained, applying separate filters to prompts and outputs on its own hosted applications, working with outside experts on data integrity screening, and supporting content authentication through watermarking and the C2PA content credentials standard, intended to help distinguish AI-generated media from human-made work. Stability AI also publishes a channel for reporting suspected misuse and describes an appeals process for accounts affected by enforcement action.
Technical ecosystem
Stability AI publishes model weights through its Hugging Face organization (over a hundred model repositories spanning image, audio, video, and 3D, with tens of thousands of followers) and inference code through its GitHub organization, which hosts over a hundred repositories including the general generative-models repo alongside more specialized ones such as stable-audio-tools for conditional audio generation and stable-virtual-camera for generative view synthesis. Hosted inference is also available through a Platform API for developers who don't want to self-host, and a self-hosted license option exists for teams that want to deploy models on their own infrastructure.
Important distinctions
Stability AI is a model developer, not a hosting community or a companion/chat product; it publishes the underlying model weights and architecture that other platforms build on. Community platforms like Civitai host derivative checkpoints and LoRAs fine-tuned from Stability AI's released models by independent creators, and those derivatives (along with how they're used) fall outside Stability AI's own acceptable use policy for its own products.
Technical context
The family dossiers separate documented architectures, licenses and compatibility boundaries behind Stability AI releases.