hamsuuu/nppe3-denoise-sr4x

NPPE-3 Low-Light Denoise + 4x Super-Resolution (RRDB, grayscale output)

RRDB backbone (5 blocks, 32 feat, grow_ch=16), scale x4. Input: RGB noisy low-res image (256x160). Output: grayscale HR luminance (1024x640) — matches what submission.py actually grades (grayscale, every 8th column).

Trained with Charbonnier loss, 5000 iterations, patch size 48, Adam + cosine LR. Leaderboard-exact val PSNR: printed in the training notebook's final cell.