Stable Cascade is the official codebase for a text-to-image generation model built on the Würstchen architecture, designed with a primary focus on computational efficiency. The repository is written primarily in Jupyter Notebook format and provides training and inference scripts alongside multiple model checkpoints. The key innovation distinguishing Stable Cascade from other diffusion models like Stable Diffusion is its use of a dramatically compressed latent space, achieving a compression factor of 42 compared to Stable Diffusion's factor of 8. This means a 1024x1024 image can be encoded to just 24x24 while maintaining crisp reconstructions, enabling significantly faster inference and cheaper training costs. Previous versions of this architecture achieved a 16x cost reduction over Stable Diffusion 1.5.
The model consists of three cascading stages. Stage A functions as a VAE with 20 million parameters and handles initial image compression. Stage B, available in 700 million and 1.5 billion parameter versions, performs additional compression as a diffusion model, with the larger variant excelling at reconstructing fine details. Stage C, the text-conditional generation model, comes in 1 billion and 3.6 billion parameter versions, with the larger version recommended due to extensive finetuning work. Despite containing 1.4 billion more parameters than Stable Diffusion XL, Stable Cascade achieves faster inference times while demonstrating superior performance in both prompt alignment and aesthetic quality according to human evaluation comparisons against Playground v2, SDXL, SDXL Turbo, and Würstchen v2.
The repository provides four primary inference notebooks covering distinct use cases. The text-to-image notebook offers basic functionality for text-to-image generation, image variation, and image-to-image operations. A dedicated ControlNet notebook demonstrates usage of provided ControlNets for inpainting, outpainting, face identity control, Canny edge detection, and super resolution tasks. A LoRA notebook shows how to finetune the text-conditional model with new tokens and LoRA layers for personalized generation. An image reconstruction notebook demonstrates the diffusion autoencoder's capability to encode images to 24x24 compressed representations and decode them back to full resolution while preserving fine details. The model is also accessible through the Hugging Face diffusers library.
Training capabilities include full model training from scratch, finetuning, ControlNet training, and LoRA training, with comprehensive documentation provided in the training folder. The repository explicitly supports known extensions including finetuning, LoRA, ControlNet, IP-Adapter, and LCM methods.
According to GitGenius activity tracking, the repository shows median issue and pull request response latency of 1569.6 hours with a mean of 1930.8 hours across twelve tracked items. The most active triagers and contributors include FurkanGozukara, brando90, and osa-mannella, each with two tracked events. The repository shares overlapping contributors with anthropics/claude-code, openhands/openhands, and huggingface/diffusers. The codebase is acknowledged to be in early development with potential for unexpected errors and non-optimized code, though the developers express commitment to releasing updates with improvements and optimizations based on community interest and contributions. All code is released under MIT LICENSE, while model weights are available under a Stability AI Non-Commercial Research Community License.
