Generative Models: GANs and VAEs

M608 Business Project in Computer Science

A comprehensive exploration of Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) across multiple datasets and architectures, including a mobile application for model deployment.

Project Overview

This project implements and compares various generative models across different datasets, focusing on:

• Deep Convolutional GANs (DCGANs) for image generation
• Conditional GANs (CGANs) for controlled generation
• Variational Autoencoders (VAEs) for latent space representation
• Cross-platform deployment through an iOS application

Project Structure

Datasets & Implementations

CIFAR-10

• DCGAN (cifar_10/dcgan/main.py)
  • Deep convolutional architecture for 32×32 color images
  • Multiple training runs with hyperparameter tuning
• VAE (cifar_10/vae/)
  • Standard VAE implementation
  • SiLU activation variant for improved stability

MNIST

• Digits DCGAN (mnist/digits/dcgan.py)
  • Classic handwritten digit generation
• Fashion MNIST (mnist/fashion/)
  • Conditional GAN for fashion items
  • VAE implementation for clothing generation

Quick Draw

• VAE (quick_draw/vae/)
  • Sketch-based generative modeling

Mobile Application

• iOS App (GANs/)
  • Swift-based mobile application
  • Real-time model inference on device
  • Integration of trained generator models

Analysis & Experiments

• analysis.ipynb - Model evaluation
• DCGAN_MNIST.ipynb - DCGAN experiments
• ipynb/vae.ipynb - VAE analysis

Model Artifacts

• models_weights/ - Saved model checkpoints from training runs
  • SafeTensors format for efficient serialization
  • Multiple snapshots across training epochs
  • Generated image samples for evaluation

Setup

Python Environment

uv sync

iOS Application

Open GANs/GANs.xcodeproj in Xcode to build and run the mobile app.

Key Features

Implemented Architectures

• DCGAN: Stable training with convolutional architectures
• Conditional GAN: Label-conditioned generation for controlled outputs
• VAE: Latent space learning with reconstruction and KL divergence losses

Training Optimizations

• Multiple output directories tracking training experiments
• Hyperparameter tuning across learning rates, batch sizes, and architectures
• Model checkpointing at regular intervals
• Loss tracking and visualization

Mobile Deployment

• Swift-MLX integration for on-device inference
• Optimized model weights for mobile performance
• Real-time generation capabilities

Project Goals

1. Comparative Analysis: Evaluate GANs vs VAEs for different generation tasks
2. Architecture Exploration: Test various network designs and activation functions
3. Practical Deployment: Deploy models in production-ready mobile application
4. Dataset Diversity: Train across multiple datasets to understand model generalization

Training Results

Model weights and generated samples are preserved in the models_weights/ directory, organized by timestamp. Each training run includes:

• Generator and discriminator/encoder-decoder weights
• Generated image samples

Future Work

• Extend to additional datasets (CelebA, ImageNet subsets)
• Implement StyleGAN architectures
• Enhance mobile app
• Compare computational efficiency across architectures

YouTube Link

Watch this video on YouTube

---

Course: M608 Business Project in Computer Science
Focus: Image Generative Models, Mobile Deployment