https://github.com/Piyushgarg06/DIV2K-SRGAN

Project Overview

This project implements a Super-Resolution GAN (SRGAN-style) using PyTorch, designed and trained under CPU-only constraints.

The goal is to upscale 32×32 low-resolution images to 64×64 high-resolution outputs, focusing on structural correctness first, followed by perceptual refinement.

The project emphasizes correct training dynamics, data consistency, and stability, rather than chasing aggressive architectures or unrealistic benchmarks.

Objectives

• Build a stable SRGAN pipeline from scratch
• Use curriculum learning to avoid GAN collapse
• Maintain correct brightness, color, and structure
• Debug and validate the full data → model → inference pipeline
• Produce reproducible, interpretable results

Dataset

• Source: DIV2K (High-Resolution images)

• Processing strategy:

  • HR images resized to 64×64
  • LR images generated by downsampling HR → 32×32

• Normalization:

  Images normalized to [-1, 1] using:

  Normalize(mean=[0.5]*3, std=[0.5]*3)
  

• Batch size: 6

• Subset used: 120 → 300 images (for feasibility on CPU)

The full dataset is excluded from version control.

A small example/ folder is included for qualitative demonstration.