A new Stable Diffusion release, namely “Stable Cascade”, has instigated significant chatter across different social platforms owing to its overnight launch. Since there are insufficient resources so far to demo or experiment with the feature, more updates are expected within a week on Automatic1111 and ComfyUI.
An interactive demo is accessible on the HuggingFace page wherein your prompt can be input, followed by clicking on ‘Run’. Once the process of GPU assignment concludes, you are presented with the result that comprises a superbly structured image and exceptionally paced text.
The interface, as evident, is user-friendly with advanced settings for customization. The ‘width’ and ‘height’ options span between 1024 and 1536 and users can create an image ratio of their preference – 1:1, 2:3, or 3:2.
As per the initial outcomes, the comprehension of prompts seems more organic, although old prompts work well too.
Turning the focus to Stable Cascade, this model is designed around the Würstchen structure. Distinguishing it from other models like Stable Diffusion, it operates with a significantly smaller latent space. This fact bears importance as a smaller latent space enhances the speed of inference and drastically reduces training cost. While Stable Diffusion involves a compression factor of 8 (a 1024×1024 image is encoded to 128×128), Stable Cascade realizes a compression factor of 42, which means a 1024×1024 image can be encoded to 24×24, ensuring sharp reconstructions. The text-conditional model has been trained in this highly compressed latent space. Compared to prior versions, this model achieves a 16x cost reduction over Stable Diffusion 1.5. Hence, its utility shines when efficiency is deemed vital. Renowned extensions such as finetuning, LoRA, ControlNet, IP-Adapter, LCM, etc., can be implemented through this method.
Several sample images have been created using the huggingface page for an initial glimpse.
Pulling together the final words, the results projected by the Stable Cascade release appear cleaner compared to SDXL, bolstering both composition and text interpretation. The understanding of natural language when prompting has also been improved. As the model begins to be supported on various user interfaces like Automatic1111 and ComfyUI, the anticipation for more tests and trials is inevitably heightened.
