Stable Cascade

Built on the Wurstchen architecture, Stable Cascade splits text-to-image generation across three stages: a 20M-parameter Stage A VAE, a Stage B diffusion decoder at 700M or 1.5B parameters, and a text-conditional Stage C at 1B or 3.6B parameters working in a latent space with a 42x compression factor that encodes 1024x1024 images down to 24x24, versus 8x for Stable Diffusion. That extreme compression makes both training and inference cheaper, since the expensive text-conditional stage operates on tiny latents while Stage B reconstructs the detail. Stability AI's repository includes four inference notebooks, LoRA training and inference, ControlNet training for inpainting, outpainting, canny edges, and super resolution, a Gradio app, and scripts for finetuning from scratch, and the models also run through Hugging Face Diffusers. Code is MIT licensed, but the model weights fall under the Stability AI Non-Commercial Research Community License, so commercial use is restricted. Released in 2024, the repository has been quiet since that summer, yet it remains a well-documented case study in highly compressed latent diffusion.