chidanandh/vision

Low-Light Denoising + 4× Super-Resolution

PyTorch checkpoint for low-light image denoising and 4× super-resolution.

Checkpoint

• File: best_sr_denoiser.pt
• Best saved epoch: 77
• Validation PSNR: 38.1631 dB
• Scale factor: 4×
• Channels: 64
• Residual blocks: 12
• LR training patch size: 64

Model

The checkpoint contains a custom PyTorch image-restoration model trained to map noisy, low-resolution RGB inputs to clean high-resolution RGB outputs.

Checkpoint dictionary keys:

• model — model state_dict
• epoch — best saved epoch
• val_psnr — validation PSNR
• config — architecture/training configuration

Loading the checkpoint

import torch

ckpt = torch.load("best_sr_denoiser.pt", map_location="cpu")
state_dict = ckpt["model"]

print("epoch:", ckpt["epoch"])
print("val_psnr:", ckpt["val_psnr"])
print("config:", ckpt["config"])

The original custom model class is required before calling load_state_dict.

Task

• Input: noisy low-light RGB image
• Output: denoised 4× super-resolved RGB image
• Optimization/evaluation target: PSNR