Hey everyone! 👋
I’m Manognya Lokesh Reddy, currently pursuing my Master’s in Artificial Intelligence at the University of Michigan-Dearborn. In this post, I’ll take you behind the scenes of one of my most technically demanding and visually exciting projects: high-resolution image synthesis using GANs.
I built a system that takes low-resolution or black-and-white historical images and transforms them into sharp, colored, high-quality outputs using a combination of AEGAN (Auto-Embedding GAN) and SRGAN (Super-Resolution GAN).
🖼️ The Problem
Low-resolution or faded historical images limit both aesthetic appeal and analytical value. Traditional upscaling doesn’t preserve fine detail, and manual enhancement is time-consuming.
The goal? Use Generative Adversarial Networks (GANs) to:
Upscale image resolution
Colorize black-and-white photos
Retain edge quality and texture
Do it fast and accurately
⚙️ Tech Stack
Python
PyTorch – for building AEGAN and SRGAN models
Docker – for reproducible deployment
JWT (JSON Web Tokens) – for secure REST API endpoints
Mixed Precision Training – to optimize GPU usage
OpenCV – for image pre/post-processing
🧠 How It Works
🔗 AEGAN Architecture
Combines auto-encoders and GANs to embed image features in a compressed space
Stabilized training with gradient penalty and Wasserstein loss
Mixed precision training helped accelerate training by ~60% on GPU
🔎 SRGAN Module
Generates 4x super-resolved outputs
Preserves perceptual detail using VGG-based content loss
Fine-tuned using perceptual + adversarial losses for photorealism
🔐 Scalable API Access
Deployed model using a Docker container
Secured API with JWT authentication to restrict access and support production scalability
📊 Results
✅ 37% improvement in FID score (Frechet Inception Distance) over baseline GANs
⚡ 60% faster inference time with GPU optimization
🖼️ Output quality good enough for archival, artistic, and medical use cases
🧪 Deployment time reduced by 50% using Dockerized workflows
💡 What I Learned
GAN training is unstable—techniques like gradient penalty and mixed precision training are critical
Embedding-based architectures like AEGAN offer superior feature reconstruction
Docker and JWT are must-have tools for secure, scalable ML deployment
Working with image data brings unique challenges (e.g., lighting variance, compression noise)
📸 Use Cases
Historical image restoration
Medical image enhancement
Satellite image upscaling
Artistic rendering for museums or documentaries
Top comments (0)