Artificial Intelligence has become one of the most popular topics in 2022. People have finally been able to try for themselves how artificial intelligence works, using randomly generated pictures from text descriptions. The most popular AIs that can do this are DALL-E and Stable Diffusion. Using them, you can create eerily realistic AI-generated images. A few days ago also released a significant update to Stable Diffusion 2. It brought a lot of improvements, but there are also disadvantages.
The most noticeable improvements are neatly summarized in Stability AI. However, they mainly touched on more accurate text prompts and realistic images. The text-to-image model is now trained with a new encoder called OpenCLIP. Also, they can now output 512×512 and 768×768 images.
Other models have also been greatly improved, such as the upscale, which can now produce much more accurate images, and the depth-image model, which can generate new images using text and an existing image. There is also an inpainting model that swaps objects around to create a new image.
However, there are some disadvantages to this update. Many users have started to complain that the new version of Stable Diffusion makes it challenging to create NSFW content and art that mimics a real artist’s style. This will result in less realistic images, so many are calling this version a “trimmed down” version. Considering that AI has been heavily criticized for being too realistic, it’s not surprising that the creators deliberately took such steps to avoid problems in the future. If you want to access the new Stable Diffusion 2, check it out on GitHub.
