4K AI Image Upscaler

PythonPyTorchReal-ESRGANComputer VisionDeep Learning

Restores and upscales low-quality images to stunning 4K resolution using Deep Learning. Trained on synthetic data to handle real-world degradations.

View Source Code

Project Overview

This project implements a state-of-the-art super-resolution model based on Real-ESRGAN architecture. The primary goal was to restore and enhance practical image quality, dealing with common degradations found in real-world images.

Key Features

•4k Upscaling: Successfully upscales low-resolution images to 4k quality with preserved details.

•Synthetic Data Training: Utilized a custom synthetic data pipeline to simulate real-world degradations like blur, noise, and compression artifacts during training.

•Optimized Inference: Fine-tuned the model for faster inference times on consumer-grade GPUs.

Technical Implementation

The system uses a Generative Adversarial Network (GAN) approach. The generator network relies on a U-Net structure with Residual-in-Residual Dense Blocks (RRDB), while the discriminator uses a U-Net architecture with spectral normalization to stabilize training.

Validation Results

Below is a comparison of the low-resolution input versus the enhanced output generated by the model.

Real-ESRGAN Comparison 1

Fig 1: Input vs Output comparison on natural scene

Real-ESRGAN Comparison 2

Fig 2: Input vs Output comparison on anime character