Dataset preparation
April 14, 2025 ยท View on GitHub
Training dataset
We utilize the widely-used DIV2K, Flickr2K, OST, FFHQ, LSDIR and DIV8K to train our enhancer. Please download them from:
| Dataset | Link |
|---|---|
| DIV2K(1-800 images for training) | Download |
| Flickr2K | Download |
| OST | Download |
| FFHQ | Download |
| LSDIR | Download |
| DIV8K | Download |
Please note that, since FFHQ is a large-scale face image dataset, we only utilize the firt $10000$ images from FFHQ.
Inference datasets preparation and how to generate DiffIQA
All of the original full-size images (which means the original images we collect from exising dataset/internet/cameras), the cropped $512 \times 512$ image patches (which serves as the standard original refence image for inferring DiffIQA dataset), and the generated DiffIQA images could be download from here,
| Dataset | Link |
|---|---|
| Full-size images | Google Drive |
| Cropped $512 \times 512$ image patches | Google Drive |
| Generated DiffIQA images (Training/Validation/Testing parts) | Google Drive |
Please note that, the cropped $512 \times 512$ image patches denote the same original reference image in DiffIQA.
Firstly, we collect 1200 images from DIV2K and Flickr2K, 1000 images from Internet under the license of Creative Commons, and we capture 640 images from real-world situations through our mobile phones or DSLR cameras. Secondly, we filter out the full-size images with large areas of flatten regions, please find the code here. Use the following command,
cd DiffEnhancer/scripts/preparation_for_diffiqa
python select_img_with_var.py --img_path [path to our provided original full-size images] --save_path [path to your save path]
Thirdly, please crop the filtered image from the 2nd step into $512 \times 512$ size, please find the code here. Use the following command,
cd DiffEnhancer/scripts/preparation_for_diffiqa
python crop_img.py --img_path [path to the filtered images from step 2nd] --save_path [path to your save path]
Finally, please generate DiffIQA through feeding the cropped images from the 3rd step into our well-trained Diffusion Enhancer. Please follow this instruction to infer DiffIQA.