Stable Diffusion Creators Unveil "Lightweight" Image Generator Stable Cascade

Results similar to Stable Diffusion XL, but generated faster.

Stability AI, the developers of Stable Diffusion, have introduced a new neural network called Stable Cascade. It utilizes the Würstchen architecture to significantly reduce the latent space. As a result, the neural network requires 16 times fewer resources during training.

In terms of results, Stable Cascade resembles Stable Diffusion XL, but it is expected to work faster. However, the authors note that the neural network may still incorrectly generate people and faces.

Among the additional features are image variation creation and prompting with an image. In the first case, the neural network experiments with a given image: it analyzes the original image and suggests similar options. Below is an example of generating variations without an additional prompt, as the neural network was simply fed the left image.

Prompting with an image works slightly differently. Noise is added to the original image, resulting in a loss of detail, after which the template is used as a basis for generation. In simple terms, the neural network creates something with the same outlines but with different details, based on the text description. In the example below, 80% noise was added to the original, along with the prompt "A person riding a rodent."

You can currently try Stable Cascade on Hugging Face - this is an unofficial online demo with limited capabilities: generation is only possible via text requests. The code is available on GitHub.

Images generated through Stable Cascade are intended for non-commercial use only.