🌧️ ClearView: Image Deraining

Remove rain streaks from images using deep learning. Upload a rainy image, pick a model, and see it transform!

Models: 5 architectures (Restormer, UNet, NAFNet Small/Mid/Large), all trained on the same mixed synthetic + real-world rain recipe (Rain13K, DDN-Data, SPA-Data, RealRain-1k-H/L). Performance: varies by model and test set, see each model's model card for full PSNR/SSIM tables. NAFNet (Large) and Restormer are the strongest performers overall; NAFNet (Small) trades some quality for a much smaller, faster model.

Runs on ZeroGPU.

📸 Try these examples:

📝 Notes:

• All 5 models share the same mixed-domain training recipe: synthetic sources (Rain13K, DDN-Data) plus real-world sources (SPA-Data, RealRain-1k-H/L)
• Works best on rain streaks resembling the training mix's density/scale, very light haze/fog is out of scope (none of these models were trained on fog)
• Real-world results vary more than synthetic benchmark results, results can differ noticeably between models on the same photo

🔗 Links:

• GitHub Repository
• Model Cards

⚠️ Limitations:

• No temporal consistency for video (processes frames independently)
• Fog/haze removal is explicitly out of scope for all 5 models (rain-only training mix)
• May lose fine details in heavily textured areas