Stable Diffusion is a deep learning, text-to-image model released in 2022. It is primarily used to generate detailed images conditioned on text descriptions, though it can also be applied to other tasks such as inpainting, outpainting, and generating image-to-image translations guided by a text prompt.[3] It was developed by the start-up Stability AI in collaboration with a number of academic researchers and non-profit organizations.[4]
Stable Diffusion is a latent diffusion model, a kind of deep generative neural network. Its code and model weights have been released publicly,[5] and it can run on most consumer hardware equipped with a modest GPU with at least 8 GB VRAM. This marked a departure from previous proprietary text-to-image models such as DALL-E and Midjourney which were accessible only via cloud services.
The development of Stable Diffusion was funded and shaped by the start-up company Stability AI.[7][8][9] The technical license for the model was released by the CompVis group at Ludwig Maximilian University of Munich.[7] 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.[4] Stability AI also credited EleutherAI and LAION (a German nonprofit which assembled the dataset on which Stable Diffusion was trained) as supporters of the project
Architecture[edit]
Stable Diffusion uses a kind of diffusion model (DM), called a latent diffusion model (LDM) developed by the CompVis group at LMU Munich.[11][5] Introduced in 2015, diffusion models are trained with the objective of removing successive applications of Gaussian noise on training images, which can be thought of as a sequence of denoising autoencoders. Stable Diffusion consists of 3 parts: the variational autoencoder (VAE), U-Net, and an optional text encoder.[12] The VAE encoder compresses the image from pixel space to a smaller dimensional latent space, capturing a more fundamental semantic meaning of the image.[11] Gaussian noise is iteratively applied to the compressed latent representation during forward diffusion.[12] The U-Net block, composed of a ResNet backbone, denoises the output from forward diffusion backwards to obtain a latent representation. Finally, the VAE decoder generates the final image by converting the representation back into pixel space.[12] The denoising step can be flexibly conditioned on a string of text, an image, or another modality. The encoded conditioning data is exposed to denoising U-Nets via a cross-attention mechanism.[12] For conditioning on text, the fixed, pretrained CLIP ViT-L/14 text encoder is used to transform text prompts to an embedding space.[5] Researchers point to increased computational efficiency for training and generation as an advantage of LDMs.[4][11]
Training data[edit]
Stable Diffusion was trained on pairs of images and captions taken from LAION-5B, a publicly available dataset derived from Common Crawl data scraped from the web, where 5 billion image-text pairs were classified based on language and filtered into separate datasets by resolution, a predicted likelihood of containing a watermark, and predicted "aesthetic" score (e.g. subjective visual quality).[13] The dataset was created by LAION, a German non-profit which receives funding from Stability AI.[13][14] The Stable Diffusion model was trained on three subsets of LAION-5B: laion2B-en, laion-high-resolution, and laion-aesthetics v2 5+.[13] A third-party analysis of the model's training data identified that out of a smaller subset of 12 million images taken from the original wider dataset used, approximately 47% of the sample size of images came from 100 different domains, with Pinterest taking up 8.5% of the subset, followed by websites such as WordPress, Blogspot, Flickr, DeviantArt and Wikimedia Commons.[15][13]
Training procedures[edit]
The model was initially trained on the laion2B-en and laion-high-resolution subsets, with the last few rounds of training done on LAION-Aesthetics v2 5+, a subset of 600 million captioned images which the LAION-Aesthetics Predictor V2 predicted that humans would, on average, give a score of at least 5 out of 10 when asked to rate how much they liked them.[16][13][17] The LAION-Aesthetics v2 5+ subset also excluded low-resolution images and images which LAION-5B-WatermarkDetection identified as carrying a watermark with greater than 80% probability.[13] Final rounds of training additionally dropped 10% of text conditioning to improve Classifier-Free Diffusion Guidance.[18]
The model was trained using 256 Nvidia A100 GPUs on Amazon Web Services for a total of 150,000 GPU-hours, at a cost of $600,000.[19][20][21]
Limitations[edit]
Stable Diffusion has issues with degradation and inaccuracies in certain scenarios. Initial releases of the model were trained on a dataset that consists of 512×512 resolution images, meaning that the quality of generated images noticeably degrades when user specifications deviate from its "expected" 512×512 resolution;[22] the version 2.0 update of the Stable Diffusion model later introduced the ability to natively generate images at 768×768 resolution.[23] Another challenge is in generating human limbs due to poor data quality of limbs in the LAION database.[24] The model is insufficiently trained to understand human limbs and faces due to the lack of representative features in the database, and prompting the model to generate images of such type can confound the model.[25]
Accessibility for individual developers can also be a problem. In order to customize the model for new use cases that are not included in the dataset, such as generating anime characters ("waifu diffusion"),[26] new data and further training are required. Fine-tuned adaptations of Stable Diffusion created through additional retraining have been used for a variety of different use-cases, from medical imaging[27] to algorithmically-generated music.[28] However, this fine-tuning process is sensitive to the quality of new data; low resolution images or different resolutions from the original data can not only fail to learn the new task but degrade the overall performance of the model. Even when the model is additionally trained on high quality images, it is difficult for individuals to run models in consumer electronics. For example, the training process for waifu-diffusion requires a minimum 30 GB of VRAM,[29] which exceeds the usual resource provided in consumer GPUs, such as Nvidia's GeForce 30 series having around 12 GB.[30]
The creators of Stable Diffusion acknowledge the potential for algorithmic bias, as the model was primarily trained on images with English descriptions.[20] As a result, generated images reinforce social biases and are from a western perspective, as the creators note that the model lacks data from other communities and cultures. The model gives more accurate results for prompts that are written in English in comparison to those written in other languages, with western or white cultures often being the default representation