KSoham/nppe_3_denoising

RRDBNet — 4x Super-Resolution / Denoising

An RRDBNet (ESRGAN-style) model trained for 4x image super-resolution and denoising, initialized from Real-ESRGAN's RealESRGAN_x4plus pretrained weights and fine-tuned on a paired low-resolution / high-resolution dataset.

Files in this repo

File	Purpose
model.py	The RRDBNet architecture (with its ResidualDenseBlock / RRDB submodules) as plain PyTorch code. Import this to reconstruct the model before loading weights.
config.json	Architecture hyperparameters used to build the model (channel counts, block count, scale factor).
model.safetensors (or best_model.pth)	Trained weights (state_dict) for the architecture above.
README.md	This file.

Architecture

• Type: RRDBNet (Residual-in-Residual Dense Block Network), the generator architecture used in ESRGAN / Real-ESRGAN.
• Feature channels: 64
• RRDB blocks: 23
• Growth channels (dense block): 32
• Upscale factor: 4x (two successive 2x nearest-neighbor upsampling stages)

Full config: see config.json.

Training details

• Initialization: pretrained on RealESRGAN_x4plus.pth, then fine-tuned.
• Loss: L1 loss between the super-resolved output and the ground-truth high-resolution image.
• Optimizer: Adam, initial LR 2e-4, cosine annealing schedule.
• Patch size: 256 (HR), with the LR crop derived at patch_size / scale_factor.
• Batch size: 8
• Epochs: 45
• Checkpoint selection: best validation PSNR across epochs.