# Stable Diffusion August 2022 Public Release In August 2022, Stability AI publicly released Stable Diffusion, a latent text-to-image model. Its open weights allowed consumer-grade GPUs to run it, marking a departure from prior proprietary tools and fueling the generative AI boom. In August 2022, Stability AI publicly released Stable Diffusion, a deep learning text-to-image model based on diffusion techniques. It is the premier product of Stability AI and is considered a key part of the ongoing [AI boom](https://www.wikiprompt.org/wiki/ai-boom). Stable Diffusion primarily generates detailed images from text descriptions but also supports tasks like [inpainting](https://www.wikiprompt.org/wiki/inpainting), outpainting, and image-to-image translation guided by a prompt. Its code and model weights were released publicly, enabling versions to run on most consumer hardware with as little as 2.4 GB VRAM - a departure from proprietary models like [DALL-E](https://www.wikiprompt.org/wiki/dall-e) and [Midjourney](https://www.wikiprompt.org/wiki/midjourney), which were accessible only via cloud services. Development involved researchers from the CompVis Group at LMU Munich and [Runway](https://www.wikiprompt.org/wiki/runway), with a computational donation from Stability AI and training data from non-profit organizations. The model is a [latent diffusion model](https://www.wikiprompt.org/wiki/latent-diffusion-model) (LDM), a type of deep generative artificial neural network. Its architecture, training procedures, and cultural impact have been widely studied. ## Development Stable Diffusion originated from a project called latent diffusion, developed in Germany by researchers at LMU Munich and Heidelberg University. Four of the original five authors - Robin Rombach, Andreas Blattmann, Patrick Esser, and Dominik Lorenz - later joined Stability AI and released subsequent versions. The technical license for the model was released by the CompVis group at LMU Munich. Development was led by Patrick Esser of Runway and Robin Rombach of CompVis, who were among the researchers who had earlier invented the latent diffusion model architecture used by Stable Diffusion. Stability AI also credited [EleutherAI](https://www.wikiprompt.org/wiki/eleutherai) and LAION, a German nonprofit that assembled the dataset on which Stable Diffusion was trained, as supporters. ## Technology ### Architecture Diffusion models, introduced in 2015, are trained to remove successive applications of Gaussian noise on training images, a sequence of denoising autoencoders. The name diffusion derives from thermodynamic diffusion, as the technique was inspired by thermodynamics. Models in the Stable Diffusion series before SD 3 used a variant called latent diffusion, developed in 2021 by the CompVis group at LMU Munich. Stable Diffusion consists of three parts: a [variational autoencoder](https://www.wikiprompt.org/wiki/variational-autoencoder) (VAE), a [U-Net](https://www.wikiprompt.org/wiki/u-net), and an optional text encoder. The VAE encoder compresses images from pixel space to a smaller dimensional [latent space](https://www.wikiprompt.org/wiki/latent-space), capturing fundamental semantic meaning. Gaussian noise is iteratively applied to the compressed latent representation during forward diffusion. The U-Net, composed of a [ResNet](https://www.wikiprompt.org/wiki/resnet) backbone, denoises the output backward to obtain a latent representation. The VAE decoder then generates the final image by converting the representation back to pixel space. The denoising step can be conditioned on text, an image, or another modality. Encoded conditioning data is exposed to denoising U-Nets via a [cross-attention](https://www.wikiprompt.org/wiki/cross-attention) mechanism. For text conditioning, a fixed, pretrained [CLIP](https://www.wikiprompt.org/wiki/clip) ViT-L/14 text encoder transforms prompts into an [embedding space](https://www.wikiprompt.org/wiki/embedding-space). Researchers cite increased computational efficiency as an advantage of LDMs. With 860 million parameters in the U-Net and 123 million in the text encoder, Stable Diffusion is relatively lightweight by 2022 standards and can run on consumer GPUs, or even CPU-only using the OpenVINO version. #### SD XL The XL version uses the same LDM architecture as previous versions but is larger, with a larger U-Net backbone, larger cross-attention context, two text encoders instead of one, and training on multiple aspect ratios (not just square). The SD XL Refiner, released concurrently, has the same architecture but was trained for adding fine details to preexisting images via text-conditional img2img. #### SD 3.0 The 3.0 version changes the backbone entirely, using a Rectified Flow Transformer instead of a U-Net, implementing the rectified flow method with a Transformer. The architecture, named "multimodal diffusion transformer" (MMDiT), has three tracks for original text encoding, transformed text encoding, and image encoding in latent space. The transformed text and image encodings are mixed during each transformer block. This differs from previous DiT versions where text encoding affected image encoding but not vice versa. ### Training Data Stable Diffusion was trained on pairs of images and captions from LAION-5B, a publicly available dataset derived from Common Crawl data. The 5 billion image-text pairs were filtered by language, resolution, predicted watermark likelihood, and aesthetic score. The dataset was created by LAION, a German non-profit receiving funding from Stability AI. The model was trained on three subsets: laion2B-en, laion-high-resolution, and laion-aesthetics v2 5+. A third-party analysis found that out of a smaller subset of 12 million images, approximately 47% came from 100 domains, with Pinterest taking 8.5%, followed by WordPress, Blogspot, Flickr, DeviantArt, and Wikimedia Commons. An investigation by Bayerischer Rundfunk showed LAION's datasets contained large amounts of private and sensitive data. ### Training Procedures The model was initially trained on laion2B-en and laion-high-resolution, with final rounds on LAION-Aesthetics v2 5+, a subset of 600 million captioned images predicted to score at least 5 out of 10 on aesthetic quality. This subset excluded low-resolution images and those with detected watermarks over 80% probability. Final rounds dropped 10% of text conditioning to improve classifier-free diffusion guidance. Training used 256 Nvidia A100 GPUs on Amazon Web Services for 150,000 GPU-hours, costing $600,000. ### Limitations Stable Diffusion has issues with degradation, particularly when generating complex compositions, hands, or text. It can exhibit biases from training data, and its tendency to replicate artistic styles has raised copyright concerns. The open nature of the model also enabled misuse for generating disinformation or non-consensual imagery, prompting discussions about content moderation and ethical deployment. ## Release and Availability The public release on August 22, 2022, via Stability AI made the model available as open weights. This contrasted with DALL-E and Midjourney, which were closed and cloud-only. Stable Diffusion could be downloaded and run locally, leading to rapid adoption and a proliferation of fine-tuned versions. An optimized version required modest GPU specs, with as little as 2.4 GB VRAM, widening access. ## Impact Stable Diffusion catalyzed public engagement with AI art, spawning communities and tools. It influenced the development of subsequent models and contributed to the AI boom. Its open weights spurred innovation but also raised legal and ethical questions about data provenance and copyright. ## Subsequent Versions After the initial release, Stability AI released improved versions including SD 2.0, SD XL, and SD 3.0. SD XL introduced architectural enhancements, while SD 3.0 adopted a transformer-based backbone. These versions continued the pattern of public releases, though changes in licensing and focus on commercial use have been noted. ## Reception and Controversy Stable Diffusion received praise for technical achievements and accessibility, but also faced criticism over training data issues, bias, and potential misuse. Legal challenges from artists and stock photo agencies emerged, questioning the legality of training on scraped images. These controversies shaped ongoing debates in generative AI policy. ## External links - [Wikipedia: Stable Diffusion](https://en.wikipedia.org/wiki/Stable_Diffusion) --- Source: https://www.wikiprompt.org/wiki/stable-diffusion-aug-2022-public-release License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-11-03T12:40:29.619272+00:00