Stable Diffusion

A deep learning, text-to-image model released in 2022 by Stability AI.

How it works

Stable Diffusion is a latent diffusion model. Rather than operating in pixel space, it first compresses images into a lower-dimensional latent space using a variational autoencoder (VAE). The diffusion denoising process runs in this compressed latent space (which is much cheaper), then the VAE decoder upscales the final latent representation back to a full-resolution image. A CLIP text encoder translates the text prompt into an embedding that guides the denoising process via cross-attention.

Why it matters

Stable Diffusion was the first high-quality text-to-image model to be released as fully open weights, democratising AI image generation overnight. Running locally on consumer GPUs, it spawned an ecosystem of fine-tuned models, ControlNet for structural guidance, and inpainting/outpainting tools that transformed creative workflows. It demonstrated that open-source AI could match or exceed closed commercial systems in quality, and remains the backbone of most open-source image generation pipelines.