Stable Diffusion is a latent text-to-image diffusion model developed through collaboration between CompVis, Stability AI, and Runway. The repository implements a generative model capable of synthesizing high-resolution images from text prompts, building on prior research in latent diffusion models published at CVPR 2022. The model uses a frozen CLIP ViT-L/14 text encoder to condition generation on text inputs, paired with an 860M UNet and 123M text encoder, making it relatively lightweight and capable of running on GPUs with at least 10GB VRAM.
The model was trained on 512x512 images from a subset of the LAION-5B database with computational support from Stability AI and LAION. Training began at 256x256 resolution before finetuning at 512x512 resolution. The repository provides multiple checkpoint versions, each representing different training stages and data filtering approaches. The sd-v1-1 checkpoint completed 237k steps at 256x256 resolution on laion2B-en followed by 194k steps at 512x512 on laion-high-resolution. Subsequent versions like sd-v1-2, sd-v1-3, and sd-v1-4 built upon earlier checkpoints with additional training on aesthetically filtered datasets and incorporated classifier-free guidance improvements.
The primary language of the repository is Jupyter Notebook, reflecting its research-oriented nature. The codebase builds heavily on OpenAI's ADM codebase and the denoising-diffusion-pytorch implementation, with transformer encoder implementation derived from the x-transformers library. The repository provides reference sampling scripts incorporating a Safety Checker Module to reduce explicit outputs and invisible watermarking to identify machine-generated images.
Beyond text-to-image generation, the repository demonstrates image modification capabilities using diffusion-denoising mechanisms inspired by SDEdit. This enables applications such as text-guided image-to-image translation and upscaling, where users can provide rough sketches or existing images that the model refines based on text prompts. The strength parameter controls the amount of noise added to input images, allowing users to balance semantic consistency with creative variation.
The model is distributed under the CreativeML OpenRAIL M license, an Open RAIL M license adapted from BigScience's work on responsible AI licensing. While commercial use is permitted, the documentation explicitly notes that the weights are research artifacts and recommends against deployment in services or products without additional safety mechanisms, given known limitations and biases present in the training data.
GitGenius tracking shows the repository has grown from 10591 to 10592 forks since July 4, 2026. Community engagement around issues and pull requests shows a median response latency of 85.8 hours across 91 tracked items, though mean latency reaches 2114.9 hours indicating some older items receive delayed attention. The most active contributors tracked by GitGenius include Vargol with 44 events, Any-Winter-4079 with 35 events, and magnusviri with 35 events. The repository shares contributors with invoke-ai/invokeai, huggingface/diffusers, and dmlc/xgboost, indicating its influence across the generative AI and machine learning ecosystem.
