README.md

---
license: cc-by-sa-4.0
task_categories:
  - image-to-image
  - image-classification
tags:
  - image-restoration
  - super-resolution
  - blind-image-super-resolution
  - clip
  - degradation
  - lpips
size_categories:
  - 100K<n<1M
pretty_name: OD-CLIP Training Dataset
---

# OD-CLIP Training Dataset

Training data for **OD-CLIP**: a degradation-aware CLIP variant that jointly
predicts degradation type and LPIPS-calibrated perceptual severity, used for
blind image super-resolution.

The dataset provides paired ground-truth (GT) and low-quality (LQ) image
crops together with per-image degradation metadata, covering four synthetic
degradation types (Gaussian blur, Gaussian noise, JPEG compression, and
downsampling) at a dense grid of physical severity levels.

## Download and Setup

The dataset is distributed as **six tar archives** (blur / jpeg_part1 /
jpeg_part2 / resize / noisy for the single-degradation set, and one tar for
the mixed-degradation set) plus two CSV manifests.
JPEG is split into two parts because its combined size exceeds Hugging Face's
per-file limit of 50 GB.

### 1. Install the CLI (once)

```bash
pip install --upgrade huggingface_hub
```

### 2. Download the dataset

```bash
# Download the whole dataset
huggingface-cli download yeeecheng/OD-CLIP \
  --repo-type dataset --local-dir OD-CLIP
```

Or download **only the parts you need**:

```bash
# Only single-degradation set (blur/jpeg/resize/noisy)
huggingface-cli download yeeecheng/OD-CLIP --repo-type dataset \
  --include "DIV2K_HR_train_crop/*" --local-dir OD-CLIP

# Only mixed-degradation set (stage-2 fine-tuning)
huggingface-cli download yeeecheng/OD-CLIP --repo-type dataset \
  --include "DIV2K_train_HR_random/*" --local-dir OD-CLIP

# Only one degradation type (e.g. blur)
huggingface-cli download yeeecheng/OD-CLIP --repo-type dataset \
  --include "DIV2K_HR_train_crop/blur.tar" \
  --include "DIV2K_HR_train_crop/merged_daclip_train_all.csv" \
  --local-dir OD-CLIP

# For JPEG, remember to grab both parts
huggingface-cli download yeeecheng/OD-CLIP --repo-type dataset \
  --include "DIV2K_HR_train_crop/jpeg_part*.tar" \
  --include "DIV2K_HR_train_crop/merged_daclip_train_all.csv" \
  --local-dir OD-CLIP
```

### 3. Extract the tar archives

```bash
cd OD-CLIP

# Extract the four type tars (single-degradation set)
cd DIV2K_HR_train_crop
for f in *.tar; do
  echo "Extracting $f ..."
  tar xf "$f"
  rm "$f"                 # optional: free disk after extraction
done
cd ..

# Note: jpeg_part1.tar and jpeg_part2.tar both extract into the same
# top level (each holds a disjoint set of jpeg{XX}.0/ subdirs), so no
# special handling is needed — the two archives merge cleanly.

# Extract the mixed-degradation tar
cd DIV2K_train_HR_random
mkdir -p train_crop
tar xf random0.tar -C train_crop/
rm random0.tar            # optional
cd ..
```

### 4. Resulting directory layout

```
OD-CLIP/
├── DIV2K_HR_train_crop/
│   ├── merged_daclip_train_all.csv          # top-level manifest
│   ├── blur0.1/  blur0.2/  ...  blur4.0/    # 40 subdirs
│   ├── jpeg31/   jpeg32/   ...  jpeg95/     # 65 subdirs
│   ├── resize1.1/  ...  resize7.0/          # 60 subdirs
│   ├── noisy1/   ...  noisy40/              # 40 subdirs
│   └── (each subdir contains: GT/, LQ/, daclip_val.csv, degraded_prompts_0.json)
└── DIV2K_train_HR_random/
    ├── merged_daclip_train_all_updated.csv  # stage-2 manifest
    └── train_crop/random0/
        ├── GT/          # 4000 clean crops
        └── LQ/          # 4000 mixed-degradation crops
```

### 5. Verify

Each single-degradation subdirectory should contain **800 GT crops and 800 LQ
crops**; the mixed-degradation `random0/{GT,LQ}` should each contain **4000**.
A quick check:

```bash
ls OD-CLIP/DIV2K_HR_train_crop/blur0.1/LQ | wc -l                # → 800
ls OD-CLIP/DIV2K_HR_train_crop/blur0.1/GT | wc -l                # → 800
ls OD-CLIP/DIV2K_train_HR_random/train_crop/random0/LQ | wc -l   # → 4000
```

---

## Contents

### 1. `DIV2K_HR_train_crop/`
Single-degradation training set. Contains 208 subdirectories, one per
`(degradation_type, level)` pair:

| Type       | Physical parameter | Range          | Step  | # subdirs |
|:-----------|:-------------------|:---------------|:------|----------:|
| Blur       | Gaussian sigma     | 0.1 → 4.0      | 0.1   | 40        |
| JPEG       | Quality factor     | 31 → 95        | 1     | 65        |
| Downsample | Scale factor       | 1.1 → 7.0      | 0.1   | 60        |
| Noise      | Gaussian sigma     | 1.0 → 40.0     | 1.0   | 40        |
|            |                    | **Total**      |       | **205**   |

*(Additional 3 clean-reference subdirs bring the total to 208.)*

Each subdirectory contains:
- `GT/` — 800 clean crops (256×256 px, PNG)
- `LQ/` — 800 degraded crops (256×256 px, PNG), paired one-to-one with `GT/`
- `daclip_val.csv` — per-image manifest (filepath, degradation label, physical
  parameter value)
- `degraded_prompts_0.json` — text prompt used for the degradation

### 2. `DIV2K_train_HR_random/`
Mixed-degradation training set. All four degradations are applied to each
image with per-image randomized parameters, used for stage-2 fine-tuning on
composite degradations.

- `train_crop/random0/GT/` — 4000 clean crops
- `train_crop/random0/LQ/` — 4000 mixed-degradation crops
- `merged_daclip_train_all_updated.csv` — top-level manifest for stage-2 training

## Source and License

Ground-truth crops are derived from the DIV2K training set
(https://data.vision.ee.ethz.ch/cvl/DIV2K/), licensed under **CC BY-SA 4.0**.
This release inherits the same license.

## Contact

For questions, please open an issue on this dataset repository.