Stable Diffusion 2.1 Balances Improved Scenery Generation with Better Portraits
Stability AI releases Stable Diffusion 2.1, offering a less aggressive NSFW filter that restores the ability to easily generate people while keeping the high-quality architectural and landscape improvements of version 2.0.
Stability AI releases Stable Diffusion version 2.1, delivering a quick update to its popular image generation model just weeks after the version 2.0 launch. This new iteration utilizes a brand new text encoder called OpenCLIP, developed by LAION, which provides a much deeper range of expression compared to the original version 1. The update supports the new prompting style introduced in version 2.0 while also bringing back many of the old prompting techniques that users enjoyed.
The development team adjusts the dataset filtering that caused issues in the previous release by making the NSFW filter less aggressive to reduce false positives. While version 2.0 significantly improved the quality of architecture, interior design, wildlife, and landscape scenes, its strict filtering unintentionally removed too many images of people. Version 2.1 fine-tunes the previous model with this adjusted filtering to capture the best of both worlds.
As a result, Stable Diffusion 2.1 easily renders beautiful architectural concepts and natural scenery while simultaneously restoring the ability to generate realistic human portraits. Users can now produce stunning portraits using familiar prompts alongside the enhanced capabilities of the updated model. The accompanying DreamStudio platform also receives updates to support these newly refined image generation features.