Quest 2 • Lesson 2
🎨 Generative AI (GANs)
Learn how Generative Adversarial Networks create new data – from images to music – by pitting two neural networks against each other.
Generative Adversarial Networks (GANs) consist of two neural networks: a Generator that creates fake data, and a Discriminator that tries to tell real from fake. They compete, improving each other over time.
"Imagine an art forger (Generator) and an art expert (Discriminator). The forger improves until the expert can't tell the difference – that's a GAN at work."
🧠 How GANs Work
- Generator – takes random noise (latent vector) and generates fake samples.
- Discriminator – classifies samples as real or fake.
- Adversarial training – generator tries to fool discriminator, discriminator tries to catch fakes.
- Loss functions – generator minimizes discriminator's ability to detect fakes; discriminator minimizes classification error.
- Applications – image generation, style transfer, text-to-image, super-resolution.
🚀 Live Demo: Train a Simple GAN
This GAN learns to generate circles from random noise. Click "Train GAN" to start. Watch the fake images become more realistic.
Real (target)
Generated (fake)
Generator Loss: —
Discriminator Loss: —
Epoch: 0
Ready. Click "Train GAN" to start.
📘 How the demo works
The GAN is built with TensorFlow.js. The generator takes 10 random numbers and outputs a 10×10 grayscale image (a circle). The discriminator tries to classify images as real or fake. After training, you can see the generator producing increasingly realistic circles.
gan.py
# Generator: noise → image
generator = Sequential([
Dense(128, activation='relu', input_shape=[10]),
Dense(100, activation='sigmoid'),
])
# Discriminator: image → real/fake
discriminator = Sequential([
Dense(64, activation='relu', input_shape=[100]),
Dense(1, activation='sigmoid'),
])
# Adversarial training loop
for epoch in range(num_epochs):
noise = random.normal(0, 1, (batch_size, 10))
fake_images = generator.predict(noise)
real_images = get_real_images()
disc_loss = discriminator.train_on_batch(real_images, ones) + discriminator.train_on_batch(fake_images, zeros)
gen_loss = gan.train_on_batch(noise, ones)
generator = Sequential([
Dense(128, activation='relu', input_shape=[10]),
Dense(100, activation='sigmoid'),
])
# Discriminator: image → real/fake
discriminator = Sequential([
Dense(64, activation='relu', input_shape=[100]),
Dense(1, activation='sigmoid'),
])
# Adversarial training loop
for epoch in range(num_epochs):
noise = random.normal(0, 1, (batch_size, 10))
fake_images = generator.predict(noise)
real_images = get_real_images()
disc_loss = discriminator.train_on_batch(real_images, ones) + discriminator.train_on_batch(fake_images, zeros)
gen_loss = gan.train_on_batch(noise, ones)
✨ Challenge: Improve the GAN
Try modifying the GAN architecture:
- Increase the latent dimension (e.g., 20 instead of 10).
- Add more layers to generator/discriminator.
- Train for more epochs.
// Generator with more layers
const generator = tf.sequential();
generator.add(tf.layers.dense({units: 64, activation: 'relu', inputShape: [20]}));
generator.add(tf.layers.dense({units: 128, activation: 'relu'}));
generator.add(tf.layers.dense({units: 100, activation: 'sigmoid'}));
const generator = tf.sequential();
generator.add(tf.layers.dense({units: 64, activation: 'relu', inputShape: [20]}));
generator.add(tf.layers.dense({units: 128, activation: 'relu'}));
generator.add(tf.layers.dense({units: 100, activation: 'sigmoid'}));
➡️ Next Lesson
Lesson 2.3: AI in Production – deploy your AI models.
Continue to Lesson 2.3 →(Coming soon – check back or buy Pro Pack for early access)
❤️ Support Free Education
This course is 100% free. If it helps you, consider buying me a coffee.
☕ Buy Me a Coffee