Stable Cascade

Stable Cascade is the latest image generation model that is 243% better than SDXL in terms of aesthetic quality, significantly better at understanding prompts and two times faster.

With Stable Cascade you can generate even more beautiful pictures with shorter prompts and shorter inference time.

Cascade Your imagination is the limit (and your GPU too)

Stable Cascade Examples

Stable Cascade

Check out the first examples, or create your own!

What is Stable Cascade

Stable Cascade is the latest AI image generation model from Stability AI based on the Würstchen architecture. The model is extremely easy to run and train on consumer-grade hardware.

One of the most significant advantages offered by Stable Cascade is its affordability in terms of training costs without compromising on quality or speed. Compared to Stable Diffusion, which compresses images from 1024×1024 to 128×128, Stable Cascade achieves a remarkable reduction, compressing the same resolution to 24×24. This results in faster inference speeds and cheaper training costs.

Stable Cascade surpasses Stable Diffusion XL by 1.4 billion parameters, promising faster inference speeds without compromising on details or quality. Its underlying structure comprises three stages: Stage A focuses on the VAE aspect with 20 million parameters; Stage B provides options between a 700 million parameter diffusion model and a more detailed 1.5 billion parameter version; Stage C offers even larger models for fine-tuning applications.

Try Stable Cascade on your own computer, on Huggingface or here on our website. See the “Stable Cascade Online Demo” section below.

Why is Stable Cascade better

Aesthetic quality

The model’s performance evaluation showed a significant increase in aesthetic quality compared to Stable Diffusion XL.

Stable Cascade performed 2.5x times better than SDXL and 5.5x times better than SDXL Turbo!

Faster inference

Due to changes in architecture, Stable Cascade is more cost effective in terms of inference resource usage. Unlike Stable Diffusion, which operates with a compression factor of 8, Stable Cascade achieves an impressive factor of 42. This allows it to compress 1024×1024 images into 24×24 dimensions without sacrificing quality and vastly improving inference speed.

Better prompt understanding

Human evaluation tests show improved prompt alignment compared to Stable Diffusion models. It means that the model understands your prompts better, both short and detailed.

Stable Cascade Online Playground

How to use the Stable Cascade

There are multiple ways that you can try this model yourself. First, you can download the model from Huggingface to your computer and run it on your own hardware. It is recommended to have about 20Gb of RAM for the biggest model.

Another option is to run it directly on our website (see section above).

Note that our demo resources are limited and you may need to try running the app several times before it generates a result.

Download Stable Cascade

Stability AI introduced the official codebase of Stable Cascade on GitHub. You can also download the training code and the weights from HuggingFace.