AI art is the hottest trend of 2022, and it’s all thanks to models like DALL-E and Stable Diffusion. Using them, you can generate eerily realistic AI-powered images. Stable Diffusion 2 has been officially released, bringing several improvements and apparently being nerfed in other aspects.
The biggest improvements in Stable Diffusion 2 have been carefully summarized by Stability AI, but basically, you can expect more accurate text prompts and more realistic images. The text-to-image models are trained with a new text encoder (OpenCLIP) and can output 512 × 512 and 768 × 768 images.
Other models are also getting a lot better, including the scaler, which can now produce much more accurate images, and the depth-to-image model, which can generate new images using both text and an existing image. There is also an internal paint model that can swap parts of an image to generate a completely new image.
However, the new update has some drawbacks. Users have complained that the new version of Stable Diffusion makes it difficult to generate NSFW content, as well as art that mimics the style of a real artist, leading some to claim that the new version has been “nerfed”. Given the heavy criticism of AI art for its ability to produce fake images that look real, it wouldn’t be surprising if the model deliberately deviates from producing images that could cause problems.
If you want access to the new Stable Diffusion 2, be sure to check out more on GitHub.
Source: AI Stability, Engadget