NPPE-3 Low-Light Denoise + 4x Super-Resolution

RRDB backbone (10 blocks, 64 features, growth channels=32), scale x4.

Input: RGB noisy low-resolution image. Output: grayscale HR luminance.

The competition score uses grayscale output and every 8th column. The model was trained with MSE loss to align optimization with PSNR.

Training:

  • iterations: 12000
  • patch size: 64
  • optimizer: Adam
  • scheduler: cosine annealing
  • learning rate: 0.0002

The final leaderboard-exact validation PSNR is printed by Cell 6.

Downloads last month
21
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support