Generative Models Quiz
Check your understanding across the course — what separates explicit from implicit density models, what the VAE's two loss terms do, how GAN training can fail, what diffusion models actually learn, and why Stable Diffusion works in latent space.
Exercise
A trained GAN and a trained VAE are both asked: 'what is p(x) for this specific image x?' What can each one do?
Need a hint?
One family models the density explicitly (or a bound on it); the other only learned to sample.
Exercise
A VAE's loss combines which two terms?
Need a hint?
One term wants faithful outputs; the other disciplines the latent space.
Exercise
What problem does the reparameterization trick solve?
Need a hint?
Can you backprop through 'draw a random sample'?
Exercise
In a GAN, what objective does the generator pursue?
Need a hint?
It never sees real data directly — only the discriminator's reaction.
Exercise
What is mode collapse in GAN training?
Need a hint?
The generator finds one thing that fools the discriminator and…
Exercise
What does a diffusion model learn to do?
Need a hint?
Training corrupts images with noise; generation runs that movie backwards.
Exercise
Stable Diffusion runs its denoising U-Net on 64×64×4 latents instead of 512×512×3 pixels. Why?
Need a hint?
Compare 786,432 values per step with 16,384.
Exercise
VAEs (ELBO), GANs (adversarial minimax), and diffusion (denoising score matching) can share nearly the same neural architecture. What is the main slot that distinguishes them?
Need a hint?
They differ mostly in what 'make realistic data' is defined to mean.
Exercise
Transfer test — a technique this course did NOT teach: gradient accumulation, summing gradients over several mini-batches before each update to simulate a larger batch. Which slot does it modify?
Need a hint?
It changes how the parameter update is computed, not the model or the loss.