ComfyUI-SwinIR

A ComfyUI custom node for SwinIR (Swin Transformer for Image Restoration) supporting image super-resolution and denoising.

Features

Multiple Model Types: Support for classical SR, lightweight SR, real-world SR, and denoising

Flexible Configuration: Customizable model parameters (window size, embed dim, depths, etc.)

Memory Efficient: Tiled processing for large images

Batch Processing: Process multiple images at once

Installation

Clone this repository into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/alexcong/ComfyUI-SwinIR.git

Install dependencies:

cd ComfyUI-SwinIR
pip install -r requirements.txt

Download SwinIR models from the official repository and place them in ComfyUI/models/upscale_models/

Usage

Nodes

SwinIR Model Loader

Loads a SwinIR model with specified configuration.

Parameters:

model_name: Select from available models in upscale_models folder

model_type: Choose model type (classicalSR, lightweightSR, realSR, denoising)

upscale: Upscale factor (1-8)

window_size: Window size for attention (default: 8)

embed_dim: Embedding dimension (default: 180)

depths: Comma-separated depths for each layer (e.g., “6, 6, 6, 6, 6, 6”)

num_heads: Comma-separated number of attention heads (e.g., “6, 6, 6, 6, 6, 6”)

mlp_ratio: MLP ratio (default: 2.0)

img_size: Training image size (default: 128) – must match model’s training size

SwinIR Upscale/Denoise

Processes images using the loaded SwinIR model.

Parameters:

swinir_model: Model from SwinIR Model Loader

images: Input images

tile_size: Tile size for processing (default: 512)

overlap: Overlap between tiles (default: 32)

Example Workflow

Add SwinIR Model Loader node

Configure model parameters to match your downloaded model

Add SwinIR Upscale/Denoise node

Connect model output to the upscale node

Connect your image input

Run!

Common Model Configurations

Classical SR (x2)

Model: 001_classicalSR_DF2K_s64w8_SwinIR-M_x2.pth

Type: classicalSR

Upscale: 2

Window Size: 8

Embed Dim: 180

Depths: “6, 6, 6, 6, 6, 6”

Num Heads: “6, 6, 6, 6, 6, 6”

MLP Ratio: 2.0

Img Size: 64 (from s64 in filename)

Lightweight SR (x2)

Model: 002_lightweightSR_DIV2K_s64w8_SwinIR-S_x2.pth

Type: lightweightSR

Upscale: 2

Window Size: 8

Embed Dim: 60

Depths: “6, 6, 6, 6”

Num Heads: “6, 6, 6, 6”

MLP Ratio: 2.0

Img Size: 64

Real-World SR (x4)

Model: 003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_GAN.pth

Type: realSR

Upscale: 4

Window Size: 8

Embed Dim: 180

Depths: “6, 6, 6, 6, 6, 6”

Num Heads: “6, 6, 6, 6, 6, 6”

MLP Ratio: 2.0

Img Size: 64

Color Denoising (Noise 25)

Model: 005_colorDN_DFWB_s128w8_SwinIR-M_noise25.pth

Type: denoising

Upscale: 1

Window Size: 8

Embed Dim: 180

Depths: “6, 6, 6, 6, 6, 6”

Num Heads: “6, 6, 6, 6, 6, 6”

MLP Ratio: 2.0

Img Size: 128 (from s128 in filename)

Testing

Run the test suite:

python test_nodes.py

Model Requirements for Testing

The test suite includes a real model loading test that requires downloading a pre-trained model:

Download 005_colorDN_DFWB_s128w8_SwinIR-M_noise25.pth from:

SwinIR GitHub Releases

Place the model in the same directory as test_nodes.py

Without the model, the real model test will be skipped. Other tests run using synthetic models and don’t require downloads.

Test Coverage

Model loading test

Basic upscaling test

Tiled processing test

Batch processing test

Real model loading test (validates the attention mask fix)

Credits

SwinIR: JingyunLiang/SwinIR

Paper: SwinIR: Image Restoration Using Swin Transformer

License

This project follows the same license as the original SwinIR repository.