Presenting Stable Cascade

Today marks the release of Stable Cascade in research preview, a novel text-to-image model built upon the Würstchen architecture, offered under a non-commercial license. This model introduces a three-stage approach which simplifies training and fine-tuning on consumer hardware. It opens up new avenues for quality, flexibility, and efficiency in image generation, reducing hardware barriers. Technical highlights include a pipeline composed of three stages, optimizing hierarchical compression of images for notable output quality with significantly reduced computational costs. Stable Cascade, compatible with inference in the diffusers library, also features added functionalities for image variation and image-to-image generation, further enhanced by the provision of code for training, fine-tuning, and specialized techniques like ControlNet and LoRA on the Stability GitHub page. This architecture not only promises high-quality outputs and new levels of efficiency but also encourages further experimentation by lowering the entry barrier for users interested in exploring this advanced technology.