Stable Cascade is an official codebase that provides training and inference scripts, as well as different models for text-to-image generation. The main difference of Stable Cascade compared to other models like Stable Diffusion is its smaller latent space, which allows for faster inference and more cost-effective training. With a compression factor of 42, Stable Cascade can encode a 1024×1024 image to 24×24 while maintaining high-quality reconstructions. The model is well-suited for efficient usage and has achieved impressive results in prompt alignment and aesthetic quality. The codebase also includes various extensions such as finetuning, ControlNet, and LoRA. Training and inference instructions are provided, with checkpoints available for each stage of the model. The codebase is undergoing development and improvements, and feedback and contributions are welcome.
https://github.com/Stability-AI/StableCascade