Image: Stable Diffusion

Stable Diffusion, developed in collaboration by the CompVis Group at Ludwig Maximilian University of Munich and other contributors, is a latent diffusion model used primarily for generating detailed images from text descriptions. Released in 2022, it differs from earlier AI models like DALL-E by being open-source and operable on consumer hardware with moderate GPU capabilities.

The model employs a technique called latent diffusion, where images are denoised in a step-wise manner from a compressed latent space representation, guided by text via a pretrained CLIP text encoder. This allows the model to generate images directly from text prompts, and it can also perform tasks like inpainting and outpainting.

Stable Diffusion was trained on the LAION-5B dataset, which includes billions of image-text pairs, sourced predominantly from English descriptions. This has led to concerns about representational bias and the reinforcement of Western-centric imagery.

Significantly, the model has sparked legal and ethical debates, particularly around the unconsented use of artists' works for training data, leading to lawsuits. Stable Diffusion's permissive approach to generated content, including potentially harmful images, has raised further concerns about its use and the responsibilities of users and developers.

Overall, Stable Diffusion represents a major advancement in accessible AI-driven image generation, though it brings with it significant challenges related to copyright, ethics, and societal impact.