Stable Diffusion 1.5

Stable Diffusion 1.5 is the text-to-image checkpoint that anchored the first wave of open image generation. It is a latent diffusion model: an 860 million parameter UNet denoises in the compressed latent space of a factor-8 autoencoder, guided by a frozen CLIP ViT-L/14 text encoder, and decodes to 512 by 512 pixel images. The v1 models were trained on subsets of the LAION dataset, including laion2B-en and an aesthetics-filtered subset, and the 1.5 checkpoint continued training from earlier versions with classifier-free guidance dropout. It was published under the CreativeML OpenRAIL-M license, which allows commercial use with usage restrictions. The linked CompVis repository contains the original research codebase and the v1 checkpoints. Although newer models surpass it in quality, SD 1.5's modest hardware needs, running on consumer GPUs with a few gigabytes of VRAM, and its enormous ecosystem of fine-tunes, LoRAs, ControlNets, and community tooling keep it in wide use among hobbyists and in products built around customized image styles.