ComfyUI-NuA-BIRD

ComfyUI implementation of "Blind Image Restoration via Fast Diffusion Inversion"</br> Original article

Features

1. Blind Deblurring
2. Non-uniform Deblurring
3. Inpainting
4. Denoising
5. Superresolution

Installation

1. Clone the repository into the ComfyUI/custom_nodes directory

   cd ComfyUI/custom_nodes
   git clone https://github.com/nuanarchy/ComfyUI-NuA-BIRD.git
   

2. Install the required modules

   pip install -r ComfyUI-NuA-BIRD/requirements.txt
   

3. Copy the model weights into the appropriate folder ComfyUI/models/checkpoints

Examples

In the examples folder, you will find the workflow diagrams, the JSON file with the configurations, and resulting images.

Workflow Diagrams

Blind Deblurring

<img src="examples/deblurring.png" alt="Blind Deblurring" width=auto height=auto>

Non-uniform Deblurring

<img src="examples/deblurring_non_uniform.png" alt="Non-uniform Deblurring" width=auto height=auto>

Inpainting

<img src="examples/inpainting.png" alt="Inpainting.png" width=auto height=auto>

Denoising

<img src="examples/denoising.png" alt="Denoising" width=auto height=auto>

Super Resolution

<img src="examples/super_resolution.png" alt="Super Resolution" width=auto height=auto>

Important

The results primarily depend on the pretrained model and the dataset</br> Limitations:

1. The model only works with square images at a resolution of 256x256 pixels
2. Faces must be cropped and centered in the images
3. For Super Resolution tasks, the input image resolution can be any size smaller than 256x256 pixels

If you want to overcome these limitations, you can train your own diffusion model using custom datasets.</br> You can use the OpenAI repository: improved-diffusion