GANs quiz

Question 1

What are the two networks in a GAN and what does each do?

Accepted Answer

Generator produces fake data; discriminator classifies real vs. fake

Answer

Both networks classify data

Answer

Generator classifies; discriminator generates

Answer

They share weights

Question 2

Why do GANs suffer from vanishing gradients early in training, and what fix is used?

Accepted Answer

-log D(G(z)) (non-saturating)

Answer

log(1 - D(G(z)))

Answer

MSE loss

Answer

Cross-entropy on real labels

Question 3

What is Wasserstein GAN (WGAN) and why is it better?

Accepted Answer

Uses Wasserstein distance with 1-Lipschitz critic instead of JS divergence

Answer

Uses standard cross-entropy

Answer

Has no discriminator

Answer

Uses KL divergence only

Question 4

What is a conditional GAN (cGAN)?

Accepted Answer

Conditions G and D on extra information (e.g., class label)

Answer

Has no discriminator

Answer

Uses only one network

Answer

Removes the noise input z

Question 5

What is the Inception Score (IS) for evaluating GANs?

Accepted Answer

Measures image quality (low p(y|x) entropy) and diversity (high p(y) entropy)

Answer

Measures pixel-level reconstruction error

Answer

Only measures image sharpness

Answer

Counts unique images generated

Question 6

What is the role of the latent space z in a GAN?

Accepted Answer

Low-dimensional prior distribution from which G samples to generate data

Answer

The discriminator's internal state

Answer

The training dataset compressed

Answer

A fixed random seed

Question 7

What is the difference between a GAN and a VAE?

Accepted Answer

VAE optimizes likelihood bound (blurry); GAN uses adversarial training (sharper)

Answer

VAE and GAN are the same

Answer

GAN optimizes log-likelihood

Answer

VAE produces sharper images

Question 8

What is StyleGAN and what does 'style' refer to?

Accepted Answer

Injects intermediate latent w into each layer via AdaIN for scale-specific style control

Answer

Uses a single global style vector

Answer

Has no mapping network

Answer

Only controls color not shape

Question 9

What is the training balance issue between G and D?

Accepted Answer

D too strong → vanishing G gradients; D too weak → no useful signal

Answer

Always run equal G and D steps

Answer

D always wins in a well-designed GAN

Answer

Use only one optimizer for both

Question 10

What is image-to-image translation with Pix2Pix?

Accepted Answer

Conditional GAN with U-Net generator + PatchGAN discriminator for paired image translation

Answer

Unconditional image synthesis

Answer

Uses only L1 loss

Answer

Requires no discriminator

Question 11

What is the PatchGAN discriminator?

Accepted Answer

Classifies image patches as real/fake rather than the whole image

Answer

Uses global average pooling

Answer

Only works for 64x64 images

Answer

Has no convolutional layers

Question 12

What is a BigGAN?

Accepted Answer

Large-scale class-conditional GAN with spectral norm and self-attention

Answer

A GAN for text generation

Answer

A recurrent GAN for video

Answer

Uses diffusion instead of adversarial training

Question 13

What is self-attention in the context of GANs (SAGAN)?

Accepted Answer

Adds self-attention to capture long-range spatial dependencies convolutions miss

Answer

Replaces all convolutions with attention

Answer

Has no discriminator

Answer

Only uses attention in the generator

Question 14

What causes checkerboard artifacts in GAN outputs?

Accepted Answer

Transposed conv with stride > kernel creates uneven pixel overlap

Answer

Batch normalization failure

Answer

Too many layers in G

Answer

D has too many parameters

Question 15

What is GAN inversion?

Accepted Answer

Finding z* such that G(z*) ≈ x for a real image x

Answer

Inverting G's weights

Answer

Training D in reverse

Answer

Decoding a VAE latent vector

Question 16

How does GAN differ from a simple autoencoder for generation?

Accepted Answer

Autoencoder uses reconstruction loss; GAN uses adversarial training from noise prior

Answer

They use the same loss

Answer

Autoencoders generate sharper images

Answer

GANs require paired data

Question 17

What role does batch normalization play in GAN training?

Accepted Answer

Normalizes activations to stabilize gradients; BN in D can interfere with WGAN Lipschitz

Answer

Only used in the discriminator

Answer

Batch norm is never used in GANs

Answer

Replaces adversarial loss

Question 18

What is the Jensen-Shannon divergence and its role in standard GANs?

Accepted Answer

Standard GAN minimizes JSD(p_data||p_G); JSD is constant when distributions don't overlap

Answer

Standard GAN minimizes KL divergence

Answer

Standard GAN maximizes mutual information

Answer

JSD always provides useful gradients

Question 19

What is text-to-image generation with GANs (e.g., StackGAN)?

Accepted Answer

Two-stage GAN: rough 64x64 image then 256x256 refinement, conditioned on text

Answer

Single-stage text-to-pixel mapping

Answer

Uses only L1 loss on text

Answer

Requires no discriminator

Question 20

How is GAN training different from standard supervised learning?

Accepted Answer

Minimax game with co-evolving loss; no fixed convergence criterion

Answer

Minimizes a fixed cross-entropy loss

Answer

Uses labeled real/fake pairs

Answer

Only updates the generator

Question 21

What is the 'global optimum' of a GAN?

Accepted Answer

p_G = p_data and D*(x) = 1/2 everywhere

Answer

G(z) = 0 for all z

Answer

D outputs 1 for all real samples

Answer

The discriminator loss = 0

Question 22

What is the effect of learning rate imbalance between G and D?

Accepted Answer

D LR too high → G vanishing gradients; G LR too high → oscillation

Answer

Equal LRs always work best

Answer

D should always have lower LR

Answer

Learning rate doesn't affect GAN stability

Question 23

What is semi-supervised learning with GANs?

Accepted Answer

D classifies K+1 classes (K real + 1 fake) using labeled+unlabeled data

Answer

D only distinguishes real vs fake

Answer

G generates labels

Answer

Only the generator is trained semi-supervised

Question 24

What GAN applications exist beyond image generation?

Accepted Answer

Data augmentation, drug discovery, super-resolution, anomaly detection, domain adaptation

Answer

Only image synthesis

Answer

Only used for facial recognition

Answer

Only for video generation

Question 25

What is the difference between unconditional and class-conditional GAN generation?

Accepted Answer

Conditional: generator takes class label y as input; unconditional: no class control

Answer

They are identical in architecture

Answer

Unconditional has better diversity

Answer

Only unconditional GANs use discriminators

Test yourself