NAFNet-SR: Low-Light Denoising + 4x Super-Resolution

Course: Deep Learning Practice (DLP 26T2) - NPPE-3
Roll No: 23F3004148
Architecture: NAFNet-SR (Nonlinear Activation Free Network with PixelShuffle 4x Head)
Best Validation PSNR: 39.38 dB

Model Overview

This model is designed for joint low-light image denoising and 4x super-resolution, recovering clean 1024x640 high-resolution images from noisy 256x160 low-resolution inputs.

Parameters: ~8.6M
Backbone: NAFNet restoration blocks with SimpleGate and channel attention (squeeze-and-excitation).
Upsampler: Two-stage PixelShuffle x4 reconstruction head with global bilinear residual skip connection.
Optimization: L1 Loss, Cosine Annealing learning rate schedule with warmup, and Exponential Moving Average (EMA).
Inference: 8-way Test-Time Augmentation (TTA).

Files

nafsr_best.pt: Best checkpoint weights.
nafsr_ema.pt: EMA smoothed model weights.

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