🌧️ ClearView: Neural Image Deraining

---

🚀 Quick Start

Try on Hugging Face

👉 Live Demo on HuggingFace

Install & Run Locally

git clone https://github.com/dronefreak/clearview.git
cd clearview
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt && pip install -e .

Inference

clearview-inference handles a single image, a directory of images, or a video, matching input/output type automatically.

Pull the latest weights from HF

from huggingface_hub import hf_hub_download

weights = hf_hub_download(
    repo_id="dronefreak/clearview-derain-unet", filename="clearview-derain-unet.pth"
)

Single image

clearview-inference --model unet --weights clearview-derain-unet.pth \
  --input rainy.jpg --output derained.jpg

Directory of images

clearview-inference --model unet --weights clearview-derain-unet.pth \
  --input-dir ./rainy_photos --output-dir ./derained_photos

Video (deraining runs frame by frame, with no temporal-consistency term, so some frame-to-frame flicker could be expected)

clearview-inference --model unet --weights clearview-derain-unet.pth \
  --input rainy_clip.mp4 --output derained_clip.mp4

---

🌍 Mixed-Domain Training

Training on a blended synthetic + real-world rain set (Rain13K, DDN-Data, SPA-Data, RealRain-1k-H/L, mildly oversampling the real-world sources) and selecting checkpoints against a blended validation metric across four of those sources, rather than optimizing for one benchmark, for a model that holds up reasonably across domains instead of maxing out a single dataset's quirks. See configs/mix/ for the exact recipe.

Training Mix

--mix-config combines 5 sources into one training set, oversampling the real-world sources 2x so the small RealRain-1k tracks aren't drowned out by the much larger synthetic sets:

Source	Type	Weight	Pairs
Rain13K	Synthetic	1.0	13,711
DDN-Data / Rain1400	Synthetic	1.0	12,600
SPA-Data	Real-world	2.0	6,385
RealRain-1k-H	Real-world	2.0	784
RealRain-1k-L	Real-world	2.0	784
Total			34,264

~77% synthetic / ~23% real by raw pair count. After the 2x real-world oversampling weight is applied (i.e. what the sampler actually draws from per epoch), that shifts to ~62% synthetic / ~38% real.

Validation / Checkpoint Selection

--val-mix-config blends 4 validation sources into one checkpoint-selection metric, so "best" means "doesn't fail badly anywhere" rather than "maxes out one dataset's quirks." A single deterministic pass, no oversampling, and SPA-Data val is capped well below its full size so it can't dominate the blended average on its own:

Source	Pairs used	Notes
SPA-Data (val split)	150	Capped from 1,000 pairs
RealRain-1k-H (validation split)	112	Full split
RealRain-1k-L (validation split)	112	Full split
Rain100L (test split)	100	Synthetic sanity anchor
Total	474

Models

10 architectures trained or wired in so far. All 9 ClearView-trained models below share the identical mixed-domain recipe, training data (5-source mix, see above) and checkpoint selection (blended validation set, see above), only batch size/accumulation steps and architecture-specific hyperparameters vary by size. Histoformer is the one exception, see the note below the table.

Model	Params	HF Model Card
Restormer [9]	15.3M	🤗 dronefreak/clearview-derain-restormer
Restormer-Small [9]	2.3M	🤗 dronefreak/clearview-derain-restormer-small
UNet (Vanilla) [10]	21.5M	🤗 dronefreak/clearview-derain-unet
NAFNet (Small) [11]	1.1M	🤗 dronefreak/clearview-derain-nafnet-small
NAFNet (Mid) [11]	14.3M	🤗 dronefreak/clearview-derain-nafnet
NAFNet (Large) [11]	116M	🤗 dronefreak/clearview-derain-nafnet-large
ResNet18-UNet [12]	14.4M	🤗 dronefreak/clearview-derain-resnet18-unet
ResNet34-UNet [12]	24.5M	🤗 dronefreak/clearview-derain-resnet34-unet
ResNet50-UNet [12]	73.3M	🤗 dronefreak/clearview-derain-resnet50-unet
Histoformer [8]	16.6M	🤗 dronefreak/Histoformer

Histoformer is wired in for inference and cross-domain comparison only. Its weights are the original authors' own all-weather (rain/raindrop/snow) checkpoint, not a ClearView training run, trained on different data, selected against a different validation set, so there is no clearview-train recipe for it here. To train Histoformer from scratch, use the official repository.

In our own spot checks (both net_g_real and net_g_best), its visible effect leaned much closer to dehazing/contrast correction than streak removal, on images with a haze or veiling component it cleaned up dramatically, but on images with clear rain streaks and no haze, including a genuine rain photograph, streaks were left largely untouched. Results can vary a lot by input, treat it as a reference point rather than a strong deraining baseline.

Benchmark Results (PSNR / SSIM)

Test Set	Domain	Restormer	NAFNet (Large)	NAFNet (Mid)	UNet (Vanilla)	ResNet50-UNet
Rain100L [1]	Synthetic	36.04 / 0.969	34.59 / 0.961	34.14 / 0.957	30.96 / 0.932	29.79 / 0.906
Rain100H [1]	Synthetic	27.78 / 0.868	27.65 / 0.856	27.72 / 0.849	26.41 / 0.823	25.37 / 0.794
Test100 [2]	Synthetic	28.65 / 0.881	27.71 / 0.865	27.96 / 0.873	24.91 / 0.836	26.16 / 0.839
Test1200 [3]	Synthetic	31.91 / 0.906	31.37 / 0.898	31.28 / 0.898	29.08 / 0.868	28.44 / 0.856
Test2800 [4]	Synthetic	32.05 / 0.928	31.75 / 0.924	31.66 / 0.923	30.61 / 0.909	28.67 / 0.883
DDN-Data [4]	Synthetic	32.21 / 0.931	31.90 / 0.927	31.84 / 0.926	30.67 / 0.912	28.72 / 0.886
SPA-Data [5]	Real-world	44.67 / 0.989	41.99 / 0.986	41.77 / 0.986	39.01 / 0.980	37.07 / 0.973
RealRain-1k-H [6]	Real-world	40.28 / 0.985	39.34 / 0.982	38.68 / 0.980	35.98 / 0.971	34.94 / 0.970
RealRain-1k-L [6]	Real-world	42.35 / 0.989	41.17 / 0.987	40.64 / 0.986	38.04 / 0.980	36.52 / 0.978
AllWeather (rain+fog) [7]	Cross-domain (stress)	13.72 / 0.584	13.53 / 0.576	13.64 / 0.579	13.66 / 0.570	13.61 / 0.555

Columns sorted highest to lowest by average PSNR across the 9 rain-only test sets (Restormer 35.10, NAFNet Large 34.16, NAFNet Mid 33.97, UNet 31.74, ResNet50-UNet 30.63). This is a curated subset, not the full model zoo, NAFNet (Small), ResNet18/34-UNet, and Histoformer are trained/evaluated too but omitted here for brevity, see their individual model cards (above) for full numbers. Metrics computed on each source's own held-out test/eval split (not the blended validation set used for checkpoint selection), full per-image distributions and logs live under runs/<model>/eval/<dataset>/.

---

📚 Supported Datasets

• Rain13K (composite synthetic set; includes Rain100H/L, Test100, Test1200, Test2800): 13.7K train pairs
• DDN-Data / Rain1400: 12.6K train / 1.4K test
• SPA-Data: real-world, video-derived rain/clean pairs
• RealRain-1k-H/L: real-world, heavy/light density tracks
• Custom: Organize as train/{rainy_image,ground_truth}, or combine any of the above via --mix-config

---

🏋️ Training and Evaluation

Training

Launch a single-GPU training run with an example command like the one below:

clearview-train \
  --model nafnet_small \
  --mix-config configs/mix/rain_mixed_synthetic_real.yaml --mix-sampler \
  --val-mix-config configs/mix/rain_mixed_val.yaml \
  --data-dir /path/to/mixed_datasets \
  --loss custom --loss-config '{"charbonnier": {"weight": 1.0}}' \
  --crop-size 256 --batch-size 64 --val-batch-size 64 --num-workers 8 \
  --optimizer adamw --lr 1e-4 --scheduler cosine --warmup-epochs 5 \
  --epochs 100 --early-stopping --patience 15 \
  --checkpoint-monitor val_psnr --checkpoint-mode max \
  --mixed-precision --ema --ema-decay 0.999 --compile \
  --output-dir ./runs/rain_mixed_nafnet_small \
  --device cuda

Evaluation

Once training finishes, evaluate a checkpoint against a labeled test set:

clearview-evaluate \
  --model unet --weights ./runs/rain_mixed_unet/checkpoints/best_val_psnr.pth \
  --data-dir /path/to/rainy_image_dataset \
  --dataset-type rain1400 \
  --rainy-dir testing/rainy_image --clean-dir testing/ground_truth \
  --batch-size 16 \
  --output-dir ./runs/rain_mixed_unet/test_eval

---

🔮 Roadmap

• Real-world rain dataset (SPA-Data, RealRain-1k-H/L, blended with synthetic sources via --mix-config)
• Add UResNet model support
• Add Restormer model support
• Add NAFNet model support
• Add Histoformer model support (inference-only)
• Mixed-domain training across the full model zoo (Restormer, UNet, NAFNet Small/Mid/Large)
• Release trained checkpoints on Hugging Face
• Temporal consistency for video
• Mobile deployment (ONNX/TensorRT)
• Snow/fog/haze removal

---

🤝 Contribute

PRs welcome! See CONTRIBUTING.md. Need help? Open an Issue.

---

📖 Citation

@software{saksena2025clearview,
  author = {Saksena, Saumya Kumaar},
  title = {ClearView: Practical Image Deraining with PyTorch},
  year = {2025},
  url = {https://github.com/dronefreak/clearview}
}

License: Apache 2.0 Author: Saumya Kumaar Saksena (@dronefreak)

References

1. Yang et al., Deep Joint Rain Detection and Removal from a Single Image, CVPR 2017 (Rain100H/L).
2. Zhang & Patel, Density-aware Single Image De-raining using a Multi-stream Dense Network, CVPR 2018 (Test100).
3. Zhang, Sindagi & Patel, Image De-raining Using a Conditional Generative Adversarial Network, IEEE TCSVT 2019 (Test1200).
4. Fu et al., Removing Rain from Single Images via a Deep Detail Network, CVPR 2017 (Test2800 / DDN-Data / Rain1400).
5. Wang et al., Spatial Attentive Single-Image Deraining with a High Quality Real Rain Dataset, CVPR 2019 (SPA-Data).
6. Li et al., RealRain-1k: A Large-Scale Dataset for Real-World Single Image Deraining, arXiv:2206.05514, 2022.
7. Li et al., Heavy Rain Image Restoration: Integrating Physics Model and Conditional Adversarial Learning, CVPR 2019 (AllWeather rain+fog / Outdoor-Rain).
8. Sun, Ren, Gao, Wang & Cao, Restoring Images in Adverse Weather Conditions via Histogram Transformer, ECCV 2024, arXiv:2407.10172 (Histoformer, inference-only baseline).
9. Zamir et al., Restormer: Efficient Transformer for High-Resolution Image Restoration, CVPR 2022, arXiv:2111.09881.
10. Ronneberger, Fischer & Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, MICCAI 2015, arXiv:1505.04597.
11. Chen, Chu, Zhang & Sun, Simple Baselines for Image Restoration, ECCV 2022, arXiv:2204.04676 (NAFNet).
12. He, Zhang, Ren & Sun, Deep Residual Learning for Image Recognition, CVPR 2016, arXiv:1512.03385 (ResNet, used as the encoder backbone for ResNet34-UNet).

---