Run Stable Cascade in Automatic1111/Forge/ComfyUI

Stable cascade is based on the Würstchen architecture, under a non-commercial license developed by StabilityAI. You can use your Stable Diffusion WebUI by Installing Stable Cascade as an extension. 

This model works on the three-layered methodology means it uses three models Stage A (VAE), Stage B, and Stage C (both are diffusion models). The first two are used to compress the image and the later one is for generating the image as 24 by 24 latent results in higher resolution.

Requirements:

• Windows Operating System
• NVIDIA Graphics Card
• 20GB or More VRAM
• Minimum 80GB of NVME or SSD storage (HDD not suitable)

Comparing Stable Cascade/SDXL/PlaygroundV2/SDXL Turbo:

The testing has been done with these four popular models with a batch size of 4 in terms of inference time. As you can see in the illustrated graph the results are quite incredible.

Stable Cascade has 1.4 billion parameters however it generates with faster inference time as compared to SDXL. The results with Stable Cascade are quite impressive as compared to SDXL.

Now, the model is so large and not optimized so, you need the medium-level version to get it used. For running this model we are using NVIDIA RTX4070 16GB VRAM.

We will do some comparison as well with Stable Cascade and SDXL but let's see how to do the installation on various Stable Diffusion WebUIs.

Installing on Automatic1111/Forge:

Currently Automatic1111/Forge officially hasn't released its extension.

1. Move to the unofficial extension for Automatic1111/Forge on GitHub. Copy the GitHub repo link (also provided below):

https://github.com/blue-pen5805/sdweb-easy-stablecascade-diffusers.git

2. Open Automatic1111 and move to the extension tab, select "Install from URL" and paste the link on the first input box.

3. At last Click "Install".

4. Restart your Automatic1111/Forge by clicking on "Apply and restart UI".

5. Once done, you will get a new tab as "Stable Cascade". Select that for generating images by putting positive/negative prompts. You can set the CFG scale, width height, steps batch size, etc.

Installing on ComfyUI:

Method 1:

If are unable to figure out what to install then use the Github repo, with this one-click method.

1. First go to the checkpoints from the Hugging Face by StabiltyAI :

https://huggingface.co/stabilityai/stable-cascade/tree/main/comfyui_checkpoints

2. Download all the checkpoints provided and put them into your "ComfyUIPortable/ComfyUI/models/checkpoints".(Your path can be different)

3. Go back to the ComfyUI into the Manager section, Update ComfyUI, and restart it to take effect.

Method 2:

1. Go to the hugging face repo link provided below  and move to the "Files and version" section:
https://huggingface.co/stabilityai/stable-cascade

2. Here, for using Stable cascade medium or high-level GPUs will going to work properly. The standard version is of a higher size. 

If you have lower VRAM then you have to choose the lite version available on the repository which is 2-4GB. Download all Stage models mentioned on the official page.

3. Move all the downloaded models into your ComfyUI_PortableFolder/ComfyUI/models folder as follows:

-Put the "Stage A" model into the "VAE" folder.

-Then Put your "Stage B" and "Stage C" models into the "UNET" folder.  

-Now, download the "model.safetensors" file into the "CLIP" folder.

4. Move to the ComfyUI into the Manager section, Update ComfyUI, and Restart it to take effect.

Testing with Stable Cascade :

We tested the model with the same prompts and settings and here are the results.

Prompt: a beautiful girl with her fingers on her face, hyper realistic, 8k, ultra detailed

After generating a lot of images we experienced that Stable Cascade performing better in detailing, images with text, and fingers. 

Yeah, we mean who is going to forget those AI weird fingers where every image model struggles. It also understanding prompts better as compared to the older diffusion models.

Now next is to deal with the typography. We have tested multiple times with earlier models like Stable Diffusion 1.5, or Stable Diffusion XL for generating typographies and they really struggle a lot. These models get confused while generating and adding text. 

Lets try some typography art with Stable Cascade.

Prompt used: "Happy Birthday" written on room wall written with melting chocolate, clear, crisp, realistic

Prompt used: "Happy Sunday" written on room wall written with melting chocolate, clear, crisp, realistic

After, testing multiple text art, we came to know that Stable Cascade can generate typography better but here is a catch. This model can generate only the normal text that is widely used like "Happy Birthday", "Happy Christmas", or something short text. If you try to generate something which is uncommon or longer that will again give you gibberish weird text.

Conclusion:

Stable Cascade has performed really well as compared to other diffusion models like Stable Diffusion 1.5 or Stable Diffusion XL. This model generates more accurate detailed and improved results with faster generation. The fined tuned model hasn't come yet but we have to wait more for optimized version.