Stable Diffusion

Stable Diffusion is a family of open-weight latent diffusion models for text-to-image generation, first introduced by Stability AI in August 2022. The series progressed through versions SD 1.x, SD 2.x, SDXL, and SD 3, with Stable Diffusion 3 released on June 12, 2024. SD 3 introduced a Multimodal Diffusion Transformer (MMDiT) architecture, a departure from the U-Net backbone used in SDXL and earlier iterations, and employs a rectified flow formulation. The model family spans 800 million to 8 billion parameters. In a technical report published by Stability AI on March 5, 2024, the 8-billion-parameter SD 3 model, using classifier-free guidance and 50 sampling steps, achieved a zero-shot FID of 24.2 on the COCO-2014 validation set at 512x512 resolution. It also recorded a 63.2% win rate against DALL-E 3 and Midjourney v6 in human preference evaluations for typography generation, as detailed in the same report. This model family matters because it represents the most prominent sustained open-weight alternative to proprietary image generation APIs, directly shaping public access to and research on generative architectures during a period of API consolidation.