> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/xinntao/Real-ESRGAN/llms.txt
> Use this file to discover all available pages before exploring further.

# Dataset Preparation

> Prepare DF2K and OST datasets for training Real-ESRGAN

Real-ESRGAN uses the DF2K (DIV2K + Flickr2K) and OST datasets for training. This guide walks through downloading and preparing these datasets.

## Download Datasets

Only high-resolution (HR) images are required. Download the following datasets:

<CardGroup cols={3}>
  <Card title="DIV2K" icon="image">
    High-quality diverse images

    [Download DIV2K\_train\_HR.zip](http://data.vision.ee.ethz.ch/cvl/DIV2K/DIV2K_train_HR.zip)
  </Card>

  <Card title="Flickr2K" icon="images">
    Flickr image collection

    [Download Flickr2K.tar](https://cv.snu.ac.kr/research/EDSR/Flickr2K.tar)
  </Card>

  <Card title="OST" icon="folder">
    OpenImages subset

    [Download OST\_dataset.zip](https://openmmlab.oss-cn-hangzhou.aliyuncs.com/datasets/OST_dataset.zip)
  </Card>
</CardGroup>

<Note>
  Extract all datasets to your `datasets/` directory. Combine DIV2K and Flickr2K into a `DF2K` folder.
</Note>

## Preparation Steps

The dataset preparation involves three steps. Steps 1 and 2 are optional but recommended for optimal training performance.

### Step 1: Generate Multi-scale Images (Optional)

For the DF2K dataset, use a multi-scale strategy to create ground-truth images at different scales. This improves the model's ability to handle various image sizes.

<Note>
  This step can be omitted for a quick trial, but is recommended for full training.
</Note>

The script generates images at multiple scales (0.75x, 0.5x, 0.33x) and a version where the shortest edge is 400 pixels:

```bash theme={null}
python scripts/generate_multiscale_DF2K.py --input datasets/DF2K/DF2K_HR --output datasets/DF2K/DF2K_multiscale
```

<Accordion title="How multi-scale generation works">
  The `generate_multiscale_DF2K.py` script:

  * Downsamples HR images to 75%, 50%, and 33% of original size
  * Creates a version with shortest edge = 400 pixels
  * Uses LANCZOS resampling for high quality
  * Saves images with suffixes like `T0`, `T1`, `T2`, `T3`

  This provides training data at multiple resolutions, helping the model generalize better.
</Accordion>

### Step 2: Crop to Sub-images (Optional)

Crop images into sub-images for faster IO and processing during training. This is especially helpful when training on slower storage.

<Note>
  Skip this step if you have fast IO or limited disk space.
</Note>

```bash theme={null}
python scripts/extract_subimages.py --input datasets/DF2K/DF2K_multiscale --output datasets/DF2K/DF2K_multiscale_sub --crop_size 400 --step 200
```

#### Parameters

<ParamField path="input" type="string" required>
  Input folder containing images to crop
</ParamField>

<ParamField path="output" type="string" required>
  Output folder for cropped sub-images
</ParamField>

<ParamField path="crop_size" type="int" default="480">
  Size of each cropped sub-image (e.g., 400x400 pixels)
</ParamField>

<ParamField path="step" type="int" default="240">
  Step size for overlapped sliding window. Smaller values create more sub-images with more overlap.
</ParamField>

<ParamField path="thresh_size" type="int" default="0">
  Minimum size threshold. Patches smaller than this are dropped.
</ParamField>

<ParamField path="n_thread" type="int" default="20">
  Number of parallel threads for processing
</ParamField>

<Accordion title="Understanding crop_size and step">
  * **crop\_size**: The dimension of each sub-image patch (e.g., 400 means 400x400)
  * **step**: How much the sliding window moves. A step of 200 with crop\_size 400 means 50% overlap
  * More overlap = more training data but requires more disk space

  Example: For a 1200x1200 image with crop\_size=400 and step=200:

  * Creates 5x5 = 25 overlapping sub-images
  * Each sub-image is 400x400 pixels
</Accordion>

### Step 3: Generate Meta Info File (Required)

Create a text file containing all image paths. This file tells the training script which images to use.

```bash theme={null}
python scripts/generate_meta_info.py --input datasets/DF2K/DF2K_HR datasets/DF2K/DF2K_multiscale --root datasets/DF2K datasets/DF2K --meta_info datasets/DF2K/meta_info/meta_info_DF2Kmultiscale.txt
```

#### Parameters

<ParamField path="input" type="list" required>
  List of input folders containing images. Can specify multiple folders separated by spaces.
</ParamField>

<ParamField path="root" type="list" required>
  Corresponding root paths for each input folder. Must have same length as `--input`.
</ParamField>

<ParamField path="meta_info" type="string" required>
  Output path for the generated meta info text file
</ParamField>

<ParamField path="check" type="boolean" default="false">
  Read each image to verify it can be loaded without errors
</ParamField>

#### Meta Info Format

The generated file contains relative paths, one per line:

```txt theme={null}
DF2K_HR_sub/000001_s001.png
DF2K_HR_sub/000001_s002.png
DF2K_HR_sub/000001_s003.png
DF2K_multiscale/000001T0.png
DF2K_multiscale/000001T1.png
...
```

<Warning>
  Each user's meta info file will be different based on how images were cropped and processed. You must generate your own file - do not use example files from the repository.
</Warning>

## Combining Multiple Datasets

You can merge multiple folders into a single meta info file:

```bash theme={null}
python scripts/generate_meta_info.py \
  --input datasets/DF2K/DF2K_multiscale_sub datasets/OST \
  --root datasets/DF2K datasets/OST \
  --meta_info datasets/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt
```

This creates a combined dataset from both DF2K and OST images.

## Validation Dataset (Optional)

If you want to validate during training, prepare a separate validation dataset with paired low-quality and high-quality images.

<Accordion title="Setting up validation">
  1. Prepare validation image pairs:
     * Ground-truth (GT) folder: High-resolution images
     * Low-quality (LQ) folder: Corresponding low-resolution images
  2. Uncomment validation sections in your training config:

  ```yml theme={null}
  val:
    name: validation
    type: PairedImageDataset
    dataroot_gt: path_to_gt
    dataroot_lq: path_to_lq
    io_backend:
      type: disk
  ```

  3. Configure validation frequency and metrics:

  ```yml theme={null}
  val:
    val_freq: !!float 5e3
    save_img: true
    metrics:
      psnr:
        type: calculate_psnr
        crop_border: 4
        test_y_channel: false
  ```
</Accordion>

## Directory Structure

After preparation, your directory structure should look like:

```
datasets/
├── DF2K/
│   ├── DF2K_HR/                    # Original HR images
│   ├── DF2K_multiscale/            # Multi-scale images
│   ├── DF2K_multiscale_sub/        # Cropped sub-images
│   └── meta_info/
│       └── meta_info_DF2Kmultiscale.txt
├── OST/                            # OST dataset
└── meta_info/
    └── meta_info_DF2Kmultiscale+OST_sub.txt
```

## Next Steps

<Card title="Train Real-ESRNet" icon="play" href="/training/train-realesrnet">
  Begin stage 1 training with your prepared dataset
</Card>