ClearView: Image Deraining with UNet

Fast and practical deep learning model for removing rain streaks from images.

Model Description

ClearView is a UNet-based image deraining model trained on 12,600 synthetic rainy images. It removes rain streaks while preserving image structure and details.

Key Features

  • Fast inference: ~0.5-1s per image on a single NVIDIA RTX 500 Ada Generation Laptop GPU
  • 🎯 Good performance: 30.9 PSNR / 0.914 SSIM on test set
  • 🔧 Easy to use: Simple PyTorch model with minimal dependencies

Performance

Metric Value Std Dev Min Max
PSNR 30.91 ±2.16 25.68 38.10
SSIM 0.914 ±0.026 0.831 0.967
MAE 0.021 ±0.006 0.009 0.040

Training Details:

  • Architecture: UNet (vanilla conv blocks)
  • Loss: L1 (pixel-wise)
  • Dataset: Rain1400 (12,600 train / 1,400 test pairs)
  • Optimizer: AdamW (lr=1e-4)
  • Training time: ~6 hours on RTX 4070 Super 12GB VRAM

Use Cases

Good for:

  • Dashcam footage enhancement
  • Surveillance video preprocessing
  • Photo restoration
  • Research baseline

⚠️ Limitations:

  • Trained on synthetic rain (may not generalize to all real-world scenarios)
  • Slight texture smoothing in heavily detailed areas
  • No temporal consistency for video currently supported (processes frames independently)

Quick Start

Gradio Demo

Try the model instantly: Space Link

Python Inference

import torch
from PIL import Image
from clearview.models import UNet

# Load model
model = UNet()
checkpoint = torch.load('clearview-unet.pth')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Load and preprocess image
image = Image.open('rainy_image.jpg')
# ... preprocessing code ...

# Inference
with torch.no_grad():
    output = model(input_tensor)

# Save result
output_image.save('derained.jpg')

Training

Trained using the ClearView repository:

clearview-train \
    --data-dir /path/to/Rain1400 \
    --model unet \
    --loss l1 \
    --optimizer adamw \
    --lr 1e-4 \
    --epochs 200 \
    --batch-size 24

Citation

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

Model Card Authors

Saumya Kumaar Saksena

Model Card Contact

GitHub Issues

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using dronefreak/clearview-derain-unet 1