Introduction to Generative Adversarial Networks with Conditions
Generative adversarial networks (GANs) are a powerful type of neural network used for generating new, synthetic data that mimics real-world data. Conditional GANs (cGANs) are a variation that allows controlling the output of a GAN by conditioning it on additional information.
How Conditional GANs Work
Like regular GANs, cGANs consist of two neural networks — a generator and a discriminator:
- The generator takes as input random noise and the condition information, and generates new data trying to fool the discriminator.
- The discriminator receives real and generated data, and the condition information, and tries to figure out which is real.
By competing against each other, both networks get better — the generator at creating more realistic fakes, and the discriminator at detecting them.
The key difference in cGANs is that both networks receive the condition information as additional input. This allows control over the output, e.g. generating images of specific classes.
Some common applications of