ComfyUI-Cog

• This is a custom node aiming to run CogView4 on diffusers while there is no official implementation on ComfyUI.
• You will need a updated version of diffusers and I don't know if updating it my break other stuff, so I advise you to make in a new instance of ComfyUI
• There is no progress bar preview in this node

Installation

Prerequisites

• ComfyUI installed and working
• GPU with at least 8GB VRAM (recommended)

Installation Steps

1. Clone this repository to your ComfyUI's custom_nodes folder:

   cd ComfyUI/custom_nodes
   git clone https://github.com/your-username/ComfyUI-Cog
   

2. Install the required dependencies:

   # For standard ComfyUI installation (non-portable):
   pip install diffusers transformers accelerate
   
   # To install the latest version of diffusers (recommended):
   pip install git+https://github.com/huggingface/diffusers.git
   
   # For ComfyUI Portable:
   .\python_embeded\python.exe -m pip install diffusers transformers accelerate
   # Or for the latest version of diffusers:
   .\python_embeded\python.exe -m pip install git+https://github.com/huggingface/diffusers.git
   

3. Restart ComfyUI

Usage

After installation, you'll have access to a new node:

CogView4 Generator

This is the main node that generates images using the CogView4 model.

Parameters:

• prompt: Text description of the image you want to generate
• width: Width of the output image (default: 1024)
• height: Height of the output image (default: 1024)
• num_inference_steps: Number of inference steps (default: 50)
• guidance_scale: Model guidance scale (default: 3.5)
• num_images: Number of images to generate per execution (default: 1)
• seed: (Optional) Seed for reproducible results

Note about Progress Bar:

The CogView4 model does not support progress callbacks, so there is no real-time progress bar available during generation. You will see console output when generation starts and finishes, but no step-by-step progress indication.

Example Workflow

1. Add the CogView4 Generator node to your workspace
2. Configure the prompt and other desired parameters
3. Connect the output of the CogView4 Generator to a Preview Image node to view the result
4. Optionally, connect to a Save Image node to save the image to disk

Performance Optimization

• Model CPU offload: loads parts of the model to CPU when not in use
• VAE slicing: processes the image in smaller slices
• VAE tiling: divides the image into blocks for processing

Troubleshooting

Import Error

If you encounter import errors like No module named 'diffusers', make sure you have installed all required dependencies.

CUDA Memory Errors

If you receive errors like CUDA out of memory, try:

• Close other applications using the GPU
• Decrease the image size parameters
• Enable the CPU offload function (enabled by default)

Slow First Run

On the first run, the model will be downloaded from Hugging Face (approximately 12GB). This can take some time depending on your internet connection.