NAFNet — Low-Light Denoising + 4x Super-Resolution

Trained from scratch on a low-light noisy image dataset (1105 train pairs, 267 val pairs). Architecture: NAFNet (Nonlinear Activation Free Network) — U-Net encoder-decoder with LayerNorm + SimpleGate blocks, PixelShuffle upsampling head, global bicubic-residual connection. Width=48, trained 60 epochs.

Validation Real PSNR: ~38 dB

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