Rain13K: Composite Single-Image Deraining Training Benchmark (Unofficial Mirror)

Unofficial redistribution of Rain13K, the composite single-image deraining training set (aggregating Rain800, Rain100H, Rain100L, Rain14000, and Rain12) used as the standard training data by MPRNet, Restormer, and most subsequent SOTA deraining papers, packaged for direct use with ClearView's dataset pipeline.

Disclaimer

This repository is not an official release of Rain13K — there isn't really an "official" single release to begin with; Rain13K is a community-assembled composite of five independently published datasets (see Dataset Sources), redistributed as a fixed bundle by the authors of MPRNet and Restormer and adopted as the de facto standard training set by dozens of subsequent papers.

This repository does not claim ownership of any images or metadata, and attributes each constituent dataset to its original creators (listed individually below). We describe them as "creators" rather than "copyright holders" deliberately — copyright in academic datasets can rest with an author's university, a funding body, or otherwise, under arrangements we have no way to verify from the outside.

On licensing — please read before relying on this mirror. We looked for a formal license for Rain13K as a whole and for each of its five constituent datasets individually, and found none in any case:

Constituent	Original paper	License found?	Notes
Rain800	Zhang et al., ID-CGAN	None	The official Google Drive host was, per the paper's own code repository, "maliciously deleted by someone" — the original source no longer exists
Rain100H / Rain100L	Yang et al., CVPR 2017	None	Original project page is now offline (404)
Rain14000	Fu et al., CVPR 2017 (= DDN-Data)	None	See our DDN-Data mirror's card for the fuller writeup
Rain12	Li et al., CVPR 2016	None	No specific red flags found either way

In the absence of an explicit license for any of these, we have chosen to preserve this widely-used composite for research reproducibility while providing full attribution and making the mirror removable on request — this is a decision we're making under that uncertainty, not a claim that we've established a legal right to redistribute it. Concretely:

• Full attribution and citation to every constituent dataset's original creators (below).
• No claim of any rights beyond hosting a copy.
• If a copyright holder, or an authorized representative, for any constituent dataset requests removal, we will comply promptly — for that specific constituent, or the whole bundle if requested.

Separately, this exact composite (same five sources, same ~13,700 pair count) has been redistributed via unrestricted Google Drive links by the authors of MPRNet and Restormer for several years, and is referenced by name in dozens of subsequent papers' code repositories, suggesting this style of redistribution is an accepted norm within this specific research community — though we recognize this doesn't establish any redistribution right on its own; the absence of an objection is not the same as permission.

If you need certainty about redistribution rights (e.g. for commercial use), contact the relevant dataset's creators or their institution directly rather than relying on this mirror.

---

Dataset Description

Rain13K is not a dataset with its own original data collection — it's a fixed composite training set assembled from five existing synthetic (and one semi-synthetic) rain datasets, standardized into a common input/target directory layout. It was popularized by MPRNet (CVPR 2021) and Restormer (CVPR 2022), and has since become the conventional training set against which nearly all subsequent SOTA single-image deraining methods report results, evaluated across five standard test benchmarks.

• Training: 13,711 rainy/clean pairs (aggregated from Rain800, Rain100H, Rain100L, Rain14000, Rain12).
• Testing (5 standard named benchmarks, evaluated separately, never mixed together):
  • Test100: 98 pairs
  • Test1200: 1,200 pairs
  • Test2800: 2,800 pairs
  • Rain100H: 100 pairs (heavy rain)
  • Rain100L: 100 pairs (light rain)

---

Changes from the Official Release

Beyond the sharding described immediately below, none: no images added, removed, or modified relative to the composite as redistributed by the MPRNet/Restormer authors. No relabeling.

Note on train/input/train/target's part_1/part_2 split: this is not a meaningful data split — it exists purely because Hugging Face Hub rejects any single directory containing more than 10,000 files, and both train/input and train/target (13,711 files each) exceed that. The two subfolders in each are an arbitrary alternating-file split of the same flat set. ClearView's Rain13KDataset (and the underlying ImagePairDataset it's built on) scans directories recursively, so pointing it at train/input/train/target (the parent of both subfolders) finds all 13,711 files transparently — no manual merging needed.

---

Dataset Structure

rain13k/
├── README.md
├── banner.jpg
├── train/
│   ├── input/
│   │   ├── part_1/     # ~6,856 images
│   │   └── part_2/     # ~6,855 images
│   └── target/
│       ├── part_1/     # ~6,856 images
│       └── part_2/     # ~6,855 images
└── test/
    ├── Test100/{input,target}      # 98 pairs each
    ├── Test1200/{input,target}     # 1,200 pairs each
    ├── Test2800/{input,target}     # 2,800 pairs each
    ├── Rain100H/{input,target}     # 100 pairs each
    └── Rain100L/{input,target}     # 100 pairs each

input/target share identical filename stems within each split (e.g. 1.jpg ↔ 1.jpg).

Split	Pairs
train	13,711
test/Test100	98
test/Test1200	1,200
test/Test2800	2,800
test/Rain100H	100
test/Rain100L	100
Total	18,009

---

Usage with ClearView

from huggingface_hub import snapshot_download
from clearview.data import Rain13KDataset, ImagePairDataset, get_train_transforms, get_val_transforms

data_dir = snapshot_download(repo_id="dronefreak/Rain13K", repo_type="dataset")

train_ds = Rain13KDataset(root_dir=data_dir, split="train", transform=get_train_transforms(crop_size=(256, 256)))

# Each test benchmark is evaluated separately (standard MPRNet/Restormer-style reporting)
rain100l_ds = ImagePairDataset(
    rainy_dir=f"{data_dir}/test/Rain100L/input",
    clean_dir=f"{data_dir}/test/Rain100L/target",
    transform=get_val_transforms(),
)

rainy, clean = train_ds[0]

Or directly via the training/evaluation CLI:

clearview-train \
  --data-dir <path-to-downloaded-snapshot> \
  --dataset-type pair \
  --train-rainy train/input --train-clean train/target \
  --val-rainy test/Rain100L/input --val-clean test/Rain100L/target \
  --model unet --batch-size 16 --crop-size 256 --epochs 100 \
  --loss l1 --mixed-precision \
  --output-dir ./runs/rain13k_unet_baseline

# Evaluate on all 5 benchmarks separately (native/variable resolution — keep
# --batch-size at its default of 1 to avoid a batch-collation crash)
for BENCH in Test100 Test1200 Test2800 Rain100H Rain100L; do
  clearview-evaluate \
    --model unet --weights ./runs/rain13k_unet_baseline/checkpoints/best_val_psnr.pth \
    --data-dir <path-to-downloaded-snapshot> \
    --dataset-type pair \
    --rainy-dir test/$BENCH/input --clean-dir test/$BENCH/target \
    --output-dir ./runs/rain13k_unet_baseline/test_eval_$BENCH
done

---

Dataset Sources

Rain13K aggregates five independently published datasets. Please cite whichever of these your evaluation protocol actually touches (all five, if training on the full composite and testing on all five benchmarks).

Rain800

Image De-raining Using a Conditional Generative Adversarial Network He Zhang, Vishwanath Sindagi, Vishal M. Patel

• arXiv: https://arxiv.org/abs/1701.05957
• Code/dataset repo: https://github.com/hezhangsprinter/ID-CGAN

Rain100H / Rain100L

Deep Joint Rain Detection and Removal from a Single Image Wenhan Yang, Robby T. Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, Shuicheng Yan

• arXiv: https://arxiv.org/abs/1609.07769
• CVPR Open Access: https://openaccess.thecvf.com/content_cvpr_2017/papers/Yang_Deep_Joint_Rain_CVPR_2017_paper.pdf

Rain14000 (= DDN-Data)

Removing Rain from Single Images via a Deep Detail Network Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, John Paisley

• See our DDN-Data mirror for full citation details (this paper has no arXiv preprint; a companion TIP 2017 paper does — arXiv:1609.02087).

Rain12

Rain Streak Removal Using Layer Priors Yu Li, Robby T. Tan, Xiaojie Guo, Jiangbo Lu, Michael S. Brown

• No arXiv preprint is available for this paper (CVPR proceedings only; a later extended version appeared in IEEE TIP without an arXiv preprint either).
• CVPR Open Access (PDF): http://www.cse.yorku.ca/~mbrown/pdf/li_cvpr16_rain.pdf
• Code: https://github.com/yu-li/LPDerain

Composite Redistribution Precedent

• MPRNet: https://github.com/swz30/MPRNet
• Restormer: https://github.com/swz30/Restormer/blob/main/Deraining/README.md

---

Attribution

All credit for the constituent datasets belongs entirely to their original creators: He Zhang, Vishwanath Sindagi, Vishal M. Patel (Rain800); Wenhan Yang, Robby T. Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, Shuicheng Yan (Rain100H/L); Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, Yinghao Liao, John Paisley (Rain14000/DDN-Data); and Yu Li, Robby T. Tan, Xiaojie Guo, Jiangbo Lu, Michael S. Brown (Rain12). Credit for assembling and popularizing this specific composite also belongs to the MPRNet and Restormer authors.

If you use this dataset in your research, please cite the relevant original publications below.

---

License

No formal license was located for Rain13K as a whole, nor for any of its five individual constituent datasets — see the Disclaimer above for our reasoning for mirroring it anyway under that uncertainty (full attribution, reproducibility purpose, removable on request). This repository makes no claim to any rights over the dataset content beyond hosting a copy.

---

Citation

If you use this dataset, please cite the constituent datasets relevant to your evaluation, plus MPRNet and/or Restormer if you're following their training/testing protocol:

@inproceedings{zhang2019image,
  title={Image de-raining using a conditional generative adversarial network},
  author={Zhang, He and Sindagi, Vishwanath and Patel, Vishal M},
  journal={IEEE transactions on circuits and systems for video technology},
  year={2019}
}

@InProceedings{Yang_2017_CVPR,
  author = {Yang, Wenhan and Tan, Robby T. and Feng, Jiashi and Liu, Jiaying and Guo, Zongming and Yan, Shuicheng},
  title = {Deep Joint Rain Detection and Removal From a Single Image},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {July},
  year = {2017}
}

@InProceedings{Fu_2017_CVPR,
  author = {Fu, Xueyang and Huang, Jiabin and Zeng, Delu and Huang, Yue and Ding, Xinghao and Paisley, John},
  title = {Removing Rain From Single Images via a Deep Detail Network},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {July},
  year = {2017},
  pages = {1715-1723}
}

@InProceedings{Li_2016_CVPR,
  author = {Li, Yu and Tan, Robby T. and Guo, Xiaojie and Lu, Jiangbo and Brown, Michael S.},
  title = {Rain Streak Removal Using Layer Priors},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2016},
  pages = {2736-2744}
}

@inproceedings{Zamir2021MPRNet,
  title={Multi-Stage Progressive Image Restoration},
  author={Zamir, Syed Waqas and Arora, Aditya and Khan, Salman and Hayat, Munawar and Khan, Fahad Shahbaz and Yang, Ming-Hsuan and Shao, Ling},
  booktitle={CVPR},
  year={2021}
}

@inproceedings{Zamir2022Restormer,
  title={Restormer: Efficient Transformer for High-Resolution Image Restoration},
  author={Zamir, Syed Waqas and Arora, Aditya and Khan, Salman and Hayat, Munawar and Khan, Fahad Shahbaz and Yang, Ming-Hsuan},
  booktitle={CVPR},
  year={2022}
}

---

Acknowledgements

We sincerely thank the creators of all five constituent datasets, and the MPRNet/Restormer authors for assembling and popularizing this composite as the field's standard training set.