What Are Generative Networks

A generative network is an artificial neural network designed to generate new data samples that resemble a given set of training data.
Unlike discriminative models, which learn to classify or predict data, generative networks learn the underlying patterns of the training data to create new, similar data points.

Depending on the domain area, there are different architectures of generative networks for music, text, and image generation. In this course, we will focus on image generation.

Image generation

The key principle of most image generation models is that we pass a random sample from the underlying distribution to the network, and it generates a full-scale image based on this sample.

What is a random sample?

A random sample in the context of generative networks refers to a set of randomly generated values drawn from a specified probability distribution, such as a Gaussian (normal) or uniform distribution. These values serve as the input or seed for the generative model, which is trained to generate valuable output from these values.

Types of generative models

There are three main architectures used to generate images:

1. Autoencoder Architecture: autoencoders utilize an encoder-decoder architecture to generate new images. The encoder compresses input data into a latent space representation, while the decoder reconstructs the original data from this representation;
2. Generative Adversarial Networks (GANs): GANs employ the generator-discriminator principle to generate new images. The generator network generates new data samples, while the discriminator network evaluates these samples to distinguish them from real data;
3. Diffusion Models: diffusion models generate new data by iteratively refining a noisy version of the target data until it closely resembles the desired output. These models start with a noisy input and gradually remove noise to generate realistic images.

Each architecture offers unique advantages and is suited for different applications in image generation.

Generated images examples

Nowadays, there are numerous pre-trained generative models capable of producing high-quality images. For instance, the images below are generated by AI: