RRDBNet — Low-Light Denoising + 4x Super-Resolution (DLP 26T2 NPPE3)

An RRDBNet (ESRGAN/Real-ESRGAN generator family) checkpoint trained end-to-end to jointly denoise low-light images and upscale them 4x, as part of the DLP 26T2 NPPE3 course competition (scored by PSNR).

Model description

Architecture: RRDBNet — a stack of Residual-in-Residual Dense Blocks (RRDB) followed by 2 pixel-shuffle upsampling stages (2x each, for a total of 4x).
Config: nf=64 feature channels, nb=23 RRDB blocks, gc=32 growth channels per dense block, scale=4.
Input / output: RGB image in [0, 1] → RGB image in [0, 1], upscaled 4x in each spatial dimension. Fully convolutional — accepts any input resolution.
Loss: Charbonnier loss (a smooth, robust L1 variant), chosen over plain L1/MSE since it tolerates residual sensor noise in the targets better.
Framework: implemented from scratch in PyTorch (no basicsr dependency), trained with a PyTorch Lightning training loop (Adam optimizer, cosine LR schedule, mixed precision, best-checkpoint selection by validation PSNR).

Intended uses & limitations

Intended for restoring low-light, noisy images from the DLP 26T2 NPPE3 dataset distribution (LR ≈ 256×160 → HR ≈ 1024×640) and similar low-light noisy photos.
Trained for a PSNR-oriented objective only — no adversarial/perceptual loss was used, so outputs prioritize pixel-accurate reconstruction over perceptual sharpness (edges may look slightly soft compared to GAN-based SR models).
Not evaluated on natural images far outside the training distribution (e.g. well-lit photos, non-photographic content, extreme upscale factors).

Training data

Paired low-resolution/noisy and high-resolution/clean images from the DLP 26T2 NPPE3 competition dataset (train/low + train/gt, validated against val/low

val/gt). Training used random 128×128 LR patches (512×512 HR) with random flip/rotation augmentation.

How to use

import torch
from rrdbnet import RRDBNet  # your model definition module

model = RRDBNet(scale=4, nf=64, nb=23, gc=32)

# Checkpoint is a PyTorch Lightning .ckpt — state dict keys are prefixed with "model."
ckpt = torch.load("best.ckpt", map_location="cpu")
state_dict = {k.removeprefix("model."): v for k, v in ckpt["state_dict"].items()}
model.load_state_dict(state_dict)
model.eval()

# inference
from PIL import Image
import torchvision.transforms.functional as TF

lr_img = Image.open("input.png").convert("RGB")
lr_tensor = TF.to_tensor(lr_img).unsqueeze(0)
with torch.no_grad():
    hr_tensor = model(lr_tensor)
hr_img = TF.to_pil_image(hr_tensor.squeeze(0).clamp(0, 1))
hr_img.save("output.png")

Evaluation

Validated by PSNR (dB) on the held-out val split of the DLP 26T2 NPPE3 dataset during training; best checkpoint selected by peak validation PSNR.

Limitations

No GAN/perceptual fine-tuning stage — if perceptual sharpness matters more than PSNR for your use case, consider fine-tuning further with an adversarial loss.
Performance on inputs with noise characteristics or lighting conditions very different from the training set is untested.

Citation

If you use this checkpoint, please cite the RRDBNet/ESRGAN architecture it's based on:

Wang, X., et al. "ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks." ECCV Workshops, 2018.

Downloads last month: -; Downloads are not tracked for this model. How to track