GAN Simulator - Interactive Deep Learning

Adversarial Game

Watch a Generator and Discriminator compete in a 2D toy world.

Dataset

Gaussian Blob
A single cluster centred at the origin, drawn from a 2-D isotropic Gaussian:
p_data(x) = 𝒩(0, 0.25·I)
i.e. each coordinate ~ 𝒩(0, σ²) with σ = 0.5.

Latent Variable

The Generator input noise vector is sampled from a Uniform distribution (not Gaussian):
z ~ Uniform(−1, 1)⁸
i.e. each of the 8 dimensions is drawn independently and uniformly from the interval [−1, 1].

Generator

Learning Rate

0.0050

Steps per Tick

Discriminator

Learning Rate

0.0050

Steps per Tick

Ready Step: 0

Live Loss Curves

💡 What to Observe

Heatmap = discriminator confidence
Red dots should move toward blue dots
Increase G steps → faster convergence
Switch datasets for different challenges

Epoch Evolution

See how the GAN improves over training epochs. Compare real MNIST with generated images.

Configuration

Training Config

Select Digit

Epoch

Epoch 1/25

💡 Key Insights

Early epochs produce noise
Loss ≠ image quality
D confusion dropping = good sign

🔵 Real Samples

🔴 Generated Samples

Diversity

50%

Sharpness

50%

D Accuracy

50%

Mode Collapse

Watch the Generator collapse to a subset of modes. Real data has 6 clusters.

Generator

Learning Rate

0.0050

Noise Dimension (dim of z)

Latent input: z ~ Uniform(−1, 1)^d
where d = Noise Dimension above. Each dimension is sampled independently and uniformly.

Techniques

Batch Normalization

Wasserstein GAN

💡 What You'll Learn

GAN may only cover one mode
High quality ≠ diversity
Higher noise dim helps
BatchNorm/WGAN reduce collapse

Mode Collapse Visualization Step: 0

Diversity Metrics

Covered Modes

Total Modes

💡 Legend

Blue dots = Real data clusters
Red dots = Generator output
Watch red dots collapse to fewer clusters
Toggle BatchNorm/WGAN and compare

Latent Explorer

Explore the continuous latent manifold. Move sliders to morph between digits smoothly.

Latent Variables

Each latent dimension z_i is sampled from
z ~ Uniform(−1, 1) — not Gaussian.
The sliders let you manually set z₁, z₂, z₃ within [−1, 1].

z₁ 0.00

z₂ 0.00

z₃ 0.00

Digit 4

💡 Key Insights

Generator learns continuous manifolds
Small z changes → smooth transitions
z₁ = digit, z₂ = morph, z₃ = variation

Interpolation Path

Click "Interpolate" to see a smooth transition between two random latent points.

Press ↔ Interpolate to generate