Stable Diffusion

• 1. Origins And Development History

  ›

  In 2021, researchers at Ludwig Maximilian University of Munich and Heidelberg University began work on a project called Latent Diffusion. Four of the original five authors later joined Stability AI to release subsequent versions of Stable Diffusion. Patrick Esser of Runway led development alongside Robin Rombach from CompVis. The team received computational donations from Stability AI and training data support from LAION, a German nonprofit organization. The model was released publicly in August 2022 with version numbers starting from 1.1 rather than 1.0. This marked a significant departure from previous proprietary text-to-image models like DALL-E and Midjourney which remained accessible only through cloud services.

  6796

• 2. Latent Diffusion Architecture Explained

  ›

  The system consists of three core components: a variational autoencoder, a U-Net block, and an optional text encoder. Gaussian noise gets iteratively applied to compressed latent representations during forward diffusion processes. The U-Net block uses a ResNet backbone to denoise outputs backwards into usable latent representations. Finally, the VAE decoder converts these representations back into pixel space for final images. With 860 million parameters in the U-Net and 123 million in the text encoder, the model runs on consumer GPUs with as little as 2.4 GB VRAM. Researchers point to increased computational efficiency as a key advantage over earlier diffusion models developed before 2021.

  101425

• 3. Evolution Of Model Versions

  ›

  Version 1.5 arrived in October 2022 with 983 million parameters initialized from weights of version 1.2. Stability AI released SD XL 1.0 in July 2023 containing 3.5 billion parameters making it approximately 3.5 times larger than previous iterations. The company introduced SD 3.0 in February 2024 featuring 800 million to 8 billion parameters across different model families. By October 2024, SD 3.5 Large reached between 2.5 billion and 8 billion parameters. Version 2.0 launched in November 2022 after being retrained from scratch on filtered datasets. Each iteration addressed specific limitations while expanding parameter counts significantly beyond initial releases.

  3233487276

• 4. Training Data And Algorithmic Bias

  ›

  Stable Diffusion trained on pairs of images and captions taken from LAION-5B, a dataset derived from Common Crawl data scraped from the web. Approximately 47% of sample images came from just 100 different domains including Pinterest which accounted for 8.5% of subsets. An investigation by Bayerischer Rundfunk revealed large amounts of private and sensitive data within LAION's datasets hosted on Hugging Face. The creators acknowledge that generated images reinforce social biases due primarily to training on English descriptions. Western or white cultures often become default representations because the model lacks sufficient data from other communities and cultures. Generated results prove more accurate for prompts written in English compared to those using other languages.

  20222430

• 5. Legal Battles And Copyright Disputes

  ›

  Three artists named Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a copyright infringement lawsuit against Stability AI in January 2023. They claimed companies infringed rights of millions of artists by training tools on five billion images scraped without consent. U.S. District Judge William Orrick inclined to dismiss most claims in July 2023 but allowed plaintiffs to file new complaints. Getty Images initiated separate legal proceedings in January 2023 alleging significant intellectual property rights violations. The High Court refused to strike out claims regarding unauthorized image usage despite Stability AI arguing all development occurred outside UK jurisdiction. As of the 4th of November 2025, Getty Images largely lost the lawsuit prompting calls for stronger copyright protections in Britain.

  8283848586

• 6. Open Source Accessibility And Ethics

  ›

  Stability AI released source code and pretrained weights publicly unlike models like DALL-E which remained proprietary. The Creative ML OpenRAIL-M license prohibits certain use cases including crime, libel, harassment, doxing, exploiting minors, giving medical advice, producing legal evidence, and discriminating against individuals based on social behavior or personal characteristics. Users own rights to generated output images and may use them commercially provided content remains legal. CEO Emad Mostaque argued that putting capabilities into public hands would result in net benefits despite potential negative consequences. A hack targeting ComfyUI extensions occurred in June 2024 with hackers claiming they targeted users who committed sins including art theft and promoting cryptocurrency. Controversy around photorealistic sexualized depictions of underage characters emerged due to such images being shared on websites like Pixiv.

  227455456