---
license: mit
tags: [super-resolution, low-light, denoising, image-restoration, pytorch]
---

# NAFNet->RRDBNet Cascade v3 - Low-Light Denoising + 4x SR

Joint low-light denoising and 4x super-resolution (256x160 -> 1024x640).
Pretrained NAFNet-SIDD-w32 denoiser cascaded into a RealESRNet-initialized
RRDBNet-23, fine-tuned end-to-end on the IITM DLP NPPE3 dataset with a
luminance-weighted Charbonnier loss and weight EMA. Leaderboard PSNR
(grayscale, stride-8): 39.888.

```python
import torch
from nafnet_cascade import CascadeModel

model = CascadeModel()
model.load_state_dict(torch.load("model_weights.pth", map_location="cpu"))
model.eval()
with torch.no_grad():
    sr = model(lr_image)  # 1x3xHxW float RGB in [0,1] -> 1x3x4Hx4W
```

Inference for the leaderboard score used x8 flip/rotation test-time augmentation.