Stable Diffusion is a brand new text-to-image & image-to-image AI generator. Huggingface is running an instance of the tool Online for the general public to use.
The Huggingface version takes a string of text & generates up to 4 images at a time. Users are offered several basic tools to control the output, such as speed of generation, number of steps, number of images & guidance scale. This Online version of Stable Diffusion takes a minute or two to generate a set of images because users need to wait in a queue & take “turns”.
Stable Diffusion was created by Stability AI & is the vision of founder & CEO Emad Mostaque. Mostaque wants to democratise AI image generation so all the code for the model is open source & the tool is free to use. Compared to other AI image generators, such as DALL-E & IMagen, Stable Diffusion not only offers users several advantages, such as the opportunity to run the pipeline on an ordinary PC, using a standard graphics card. Developers can, also, tweak the code to suit specific use cases.
One of the most controversial things about Stable Diffusion is that it does not have the same level of safety filters built into the system that other tools such as OpenAI do have. This means that the tool could, perceivable, be used for nefarious purposes such as generating deep fakes & fake news. All other AI tools have filters excluding famous profile images, graphic pornography & graphically violent images, but not Stable Diffusion—so, users can pretty much generate anything they please using the tool on their own machines. The Huggingface instance (thank goodness) does have a NSFW (Not Safe for Work) filter built into the interface.
Stability AI used LAION to collect huge datasets of images from across the Web to train Stable Diffusion. Most of the images were collected from Pinterest, but others have been sourced from hosted blogs on WordPress, Flickr, DeviantArt & Wikimedia, among others. The AI was originally trained on over 2.3 billion images, of which 12 million have been put Online (datasette.io) for the general public to view.
As to the quality of images generated by Stable Diffusion many reviewers think that the results are on a par with closed source rivals. However, judging from the explosion of images on social media Sites, like Twitter, the images are top notch, especially those generated to depict photorealistic landscapes. Other excellent images include those generated to portray fictional characters, such as Batman, The Incredible Hulk, Spiderman & Harry Potter among others. Fictional character dataset images are not included in other AI Image Learning models.
Mostaque believes that the public can be trusted to use the tool he’s built well. Many journalists & policy makers, however, are not so sure. As a result we can expect lawmakers across the globe to take these new tools seriously & produce a guide for their use – copyright eligibility & so forth – in the very near future.
All images generated using Stable Diffusion
