Swinir

SwinIR x4 Image Super-Resolution

This repository contains a fine-tuned SwinIR (Swin Transformer for Image Restoration) model for 4× image super-resolution.

The model takes a low-resolution RGB image as input and generates a 4× higher-resolution image.

Model Details

Model Architecture

The model was created using the following SwinIR configuration:

model = net(
    upscale=4,
    in_chans=3,
    img_size=64,
    window_size=8,
    img_range=1.0,
    depths=[6, 6, 6, 6, 6, 6],
    embed_dim=180,
    num_heads=[6, 6, 6, 6, 6, 6],
    mlp_ratio=2,
    upsampler="nearest+conv",
    resi_connection="1conv"
)

Training

The model was initialized from a pretrained SwinIR ×4 checkpoint and fine-tuned on paired low-resolution and ground-truth images.

Training setup

• Loss: L1 Loss
• Optimizer: AdamW
• Learning Rate: 2e-5
• Weight Decay: 1e-4
• Gradient Clipping: 1.0
• Learning Rate Scheduler: Cosine Annealing
• Training Patch: 64 × 64 LR
• Corresponding Ground Truth Patch: 256 × 256
• Data Augmentation: Random horizontal and vertical flips

Model Checkpoint

The main model file is:

best_swinir.pth

This file contains the fine-tuned SwinIR model checkpoint selected based on validation PSNR.

Usage

To use the model, first create the SwinIR architecture with the configuration described above and load the checkpoint.

Example:

import torch
import sys

sys.path.append("/path/to/SwinIR")

from models.network_swinir import SwinIR as net

model = net(
    upscale=4,
    in_chans=3,
    img_size=64,
    window_size=8,
    img_range=1.0,
    depths=[6, 6, 6, 6, 6, 6],
    embed_dim=180,
    num_heads=[6, 6, 6, 6, 6, 6],
    mlp_ratio=2,
    upsampler="nearest+conv",
    resi_connection="1conv"
)

checkpoint = torch.load(
    "best_swinir.pth",
    map_location="cpu"
)

if "model" in checkpoint:
    model.load_state_dict(checkpoint["model"])
else:
    model.load_state_dict(checkpoint)

model.eval()

Input and Output

Low-Resolution RGB Image
          ↓
        SwinIR
          ↓
4× Super-Resolved Image

For example:

64 × 64  →  256 × 256

Repository Structure

.
├── best_swinir.pth
├── config.json
└── README.md

Files

• best_swinir.pth — Fine-tuned SwinIR model checkpoint
• config.json — Model architecture configuration
• README.md — Model documentation

Intended Use

This model is intended for image super-resolution experiments where a low-resolution RGB image needs to be enhanced to a 4× higher spatial resolution.

Limitations

The model was trained and evaluated on the dataset used for the associated super-resolution competition. Performance may vary on images with characteristics significantly different from the training data.

License

This repository contains a fine-tuned model based on the SwinIR architecture. Please refer to the original SwinIR repository and its license for the underlying architecture and pretrained model.