Stability AI Releases New AI Model, Stable Cascade

Stability AI has launched Stable Cascade, a new AI model for generating images from text. This model, based on the Würstchen architecture, a text-to-image  architecture for large-scale text-to-image diffusion models. 

Central to Stable Cascade's appeal is its ability to closely adhere to user prompts, a feature highlighted by the AI community on HackerNews for its precision in generating relevant images. “This model is noted for its speed, offering a significant improvement in processing times without sacrificing output quality, addressing a critical need for efficiency in AI-driven tasks,” the users point out. 

Available for non-commercial use, it introduces a three-stage process designed to work efficiently on consumer hardware.

Stable Cascade is different from previous models because it uses a three-part system to compress and generate images. This allows for significant reductions in the resources needed for training. The model consists of stages A, B, and C, with stage C focusing on generating a compressed version of the image which is then expanded in stages A and B.

The company has also released tools for training and customising the model. These include scripts for fine-tuning and other adjustments, available on the Stability GitHub page. The model supports features like image variations and image-to-image generation, adding to its versatility.

In comparison tests, Stable Cascade showed better results in terms of speed and quality against other models, even those with more parameters. It offers various options for users, including different model sizes to accommodate different hardware capabilities.

Stability AI has made all related code public for users to modify and experiment with. This includes features for enhancing images, generating images from sketches, and increasing image resolution.

While Stable Cascade is not for commercial use, Stability AI suggests other models for those interested in commercial applications. Users on the latest