anant80/dlp-nppe3-submitted-model

library_name: pytorch tags:

• —swin2sr
• —super-resolution
• —image-restoration
• —denoising
• —low-light
• —kaggle ---

Swin2SR ×4 Fine-tuned for Low-Light Restoration

Base model:

caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr

Fine-tuned for noisy low-light image restoration and 4× super-resolution.

Results

• —Best validation mean PSNR: 39.5129 dB
• —Best validation global sampled PSNR: 39.2407 dB
• —Kaggle leaderboard score with 8-way TTA: 39.62684

Input

RGB low-resolution image.

Output

4× super-resolved RGB image. For the competition submission, predictions were converted to grayscale and sampled using flatten()[::8].