Stability AI

Stability AI builds foundational open-weights generative models for images, audio, video, and 3D assets—you can run the AI on your own servers or plug it straight into production pipelines via API. While multimodal workspaces like Luma Dream Machine or CoreMind bundle third-party engines inside a cloud browser, Stability AI actually builds the underlying architectures and lets you download the weights for full local control.

The company has grown way past a community lab: with Visual Effects pioneer Robert Legato on board and $232M in total funding from Electronic Arts, Universal Music Group, Warner Music Group, and Sony Music Group, the focus is now squarely on commercially safe enterprise production.

Here’s what is under the hood across formats:

• Images (Stable Diffusion & Stable Image): Text-to-image and image-to-image generation, inpainting, outpainting, background removal, relighting, search-and-replace without manual masking, structure control via ControlNet and sketches, plus upscaling up to 4 megapixels.
• Audio (Stable Audio): Trained on fully licensed datasets, these models generate full tracks up to 6 minutes long, ambient soundscapes, and isolated SFX. A compact version built with Arm runs offline right on mobile CPUs, while producers can drop an official DAW plugin straight into their audio workstations—a rare native integration compared to browser-first music generators like Suno.
• Video & 3D (Stable Video, Stable Fast 3D, Stable Virtual Camera): Reconstruct UV-textured 3D meshes and point clouds from a single image in seconds, generate multi-view 4D dynamic assets from video, or turn 2D pictures into 3D videos with camera path control.
• Language & Code (Stable LM & Stable Code): Compact language models for text tasks in seven European languages and code autocomplete.

For deployment, you can self-host models on your own hardware, connect via the REST/gRPC API, or launch through Amazon Bedrock, Amazon SageMaker, Microsoft Azure AI Foundry, and NVIDIA NIM microservices. Corporate teams also get Brand Studio to train custom models on brand books and product catalogs.

The licensing math is simple: under the Community License, self-hosting the core models is completely free for researchers, indie creators, and businesses making under $1M in annual revenue. You own the outputs and can freely train and sell LoRAs or fine-tunes—the only catch is you can't use the weights or outputs to train a competing foundational model architecture. Once your company crosses the $1M annual revenue mark, commercial use requires a custom Enterprise License, which adds legal indemnification and implementation support.