https://github.com/prathamesh-88/micro-gan

A simple implementation of Generative Adversarial Network (GAN) to generate previously unseen images based on training images.
https://github.com/prathamesh-88/micro-gan

Last synced: 9 days ago
JSON representation

A simple implementation of Generative Adversarial Network (GAN) to generate previously unseen images based on training images.

• Host: GitHub
• URL: https://github.com/prathamesh-88/micro-gan
• Owner: prathamesh-88
• License: mit
• Created: 2023-09-05T17:36:48.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2023-09-05T18:52:24.000Z (about 1 year ago)
• Last Synced: 2024-04-24T10:06:52.286Z (7 months ago)
• Language: Python
• Size: 625 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Micro GAN (Generative Adversarial Network)

## Generative Adversarial Networks

Generative Adversarial Network or GAN for short is a deep-learning based training architecture. It consists of 2 models (neural networks) acting as each other's adversaries that optimize each other during the training process. This architecture is responsible for most of your favourite text-to-image and image generation models like Midjourney.

## Micro GAN

Models like Midjourney, Stable Diffusion, Imagegen are highly sophisticated. They are trained on millions of images with billions of parameters. This repo showcases, a very minimal and simple example of a GAN that generates image based on the images it is trained on.

For the sake of simplicity, it generates a 64x64 pixel image. This can be changed by altering the models based on the size of image you wish to generate.

## Components

A generic GAN consists of 2 parts. 

1. A **Generator** that generates the images.

2. A **Discriminator** that differentiates between a real image and generated image  

### Generator

The architecture of the Generator model looks something like this:

![Architecture for generator](./diagrams/generator.png "Generator Model")

### Discriminator

The architecture of the Discriminator model looks something like this:

![Architecture for discriminator](./diagrams/discriminator.png "Discriminator Model")

### Dataset

For training this model, I used the Abstract art dataset from Kaggle.

Link to dataset: [https://www.kaggle.com/datasets/bryanb/abstract-art-gallery](https://www.kaggle.com/datasets/bryanb/abstract-art-gallery)