Diffusion models quiz

Question 1

In a DDPM, what does the forward (diffusion) process do, and is it learned?

Accepted Answer

A fixed Gaussian noising Markov chain with no learned parameters

Answer

A learned encoder that compresses data to latents

Answer

A learned chain that denoises from noise to data

Answer

A deterministic interpolation with trainable variances

Question 2

What is the closed-form expression for sampling x_t directly from x_0 in DDPM, and why is it useful?

Accepted Answer

x_t = sqrt(alpha_bar_t) x_0 + sqrt(1-alpha_bar_t) epsilon

Answer

x_t = alpha_bar_t x_0 + (1-alpha_bar_t) epsilon

Answer

x_t = sqrt(beta_t) x_0 + sqrt(1-beta_t) epsilon

Answer

x_t = x_0 + t*epsilon

Question 3

The simplified DDPM training objective (L_simple) trains the network to predict what quantity?

Accepted Answer

The added Gaussian noise epsilon

Answer

The clean image x_0 directly

Answer

The posterior variance

Answer

The log-likelihood gradient of x_t

Question 4

How does the DDPM/score-based denoising objective relate to score matching?

Accepted Answer

score = -epsilon_theta / sqrt(1-alpha_bar_t)

Answer

score = epsilon_theta * sqrt(alpha_bar_t)

Answer

score = x_t - epsilon_theta

Answer

score = log epsilon_theta

Question 5

What is the key difference between DDIM and DDPM sampling?

Accepted Answer

DDIM uses a non-Markovian, optionally deterministic reverse process allowing fewer steps

Answer

DDIM requires retraining with a different loss

Answer

DDIM adds more noise at each step than DDPM

Answer

DDIM cannot reproduce DDPM samples under any setting

Question 6

In classifier guidance, how is the diffusion sampling modified to condition on a class y?

Accepted Answer

Add s * grad_x log p(y|x_t) to the unconditional score

Answer

Multiply the score by the class probability

Answer

Replace the score with the classifier logits

Answer

Concatenate y to the input only, no score change

Question 7

How does classifier-free guidance (CFG) work and what is its main advantage over classifier guidance?

Accepted Answer

eps = eps_uncond + w*(eps_cond - eps_uncond)

Answer

eps = eps_cond + w*eps_uncond

Answer

eps = w*eps_cond

Answer

eps = (eps_cond + eps_uncond)/2

Question 8

What does increasing the CFG guidance scale w typically do to generated samples?

Accepted Answer

Increases prompt fidelity but lowers diversity and can oversaturate

Answer

Always improves both fidelity and diversity

Answer

Reduces fidelity while increasing diversity

Answer

Has no effect on diversity

Question 9

What is the core idea of Latent Diffusion Models (e.g., Stable Diffusion)?

Accepted Answer

Diffuse in a learned compressed latent space instead of pixel space

Answer

Diffuse directly in pixel space at higher resolution

Answer

Replace the U-Net with a transformer encoder only

Answer

Use no autoencoder, just downsampled pixels

Question 10

How is text conditioning typically injected into the U-Net of a text-to-image latent diffusion model?

Accepted Answer

Cross-attention with text embeddings as keys/values

Answer

Concatenation of text tokens to the noisy latent channels

Answer

Adding text embeddings to the timestep embedding only

Answer

A separate GAN discriminator on text

Question 11

Why is the timestep t embedded and fed into every block of the diffusion U-Net?

Accepted Answer

So the shared network adapts behavior to the current noise level

Answer

To increase parameter count

Answer

To act as positional encoding for pixels

Answer

To normalize the input image

Question 12

Why is the U-Net architecture well-suited for diffusion denoising?

Accepted Answer

Skip connections preserve detail while the bottleneck captures global context

Answer

It is the only architecture with a softmax output

Answer

It avoids convolutions entirely

Answer

It guarantees invertibility of the denoising map

Question 13

What is the role of the noise variance schedule (e.g., linear vs cosine) in diffusion models?

Accepted Answer

Cosine destroys signal more gradually, improving quality over linear

Answer

Linear always outperforms cosine

Answer

The schedule has no measurable effect

Answer

Cosine adds all noise in the first step

Question 14

In Song et al.'s score SDE framework, what do the forward and reverse processes correspond to?

Accepted Answer

Forward SDE perturbs data to noise; reverse-time SDE (using the score) generates data

Answer

Both are deterministic ODEs with no noise

Answer

Forward is learned, reverse is fixed

Answer

They describe a GAN min-max game

Question 15

What is the probability-flow ODE and why is it practically important?

Accepted Answer

A deterministic ODE matching the SDE marginals, enabling fast sampling and exact likelihood

Answer

A stochastic process with extra injected noise

Answer

An approximation only valid at t=0

Answer

A discriminator network

Question 16

What problem do consistency models solve, and what is their defining property?

Accepted Answer

Map any point on an ODE trajectory to its origin for one-step generation

Answer

Add more diffusion steps for higher quality

Answer

Train a classifier on noisy data

Answer

Replace the U-Net with an RNN

Question 17

What is the central idea of flow matching / rectified flow, and how does it relate to diffusion?

Accepted Answer

Regress a velocity field transporting noise to data, often along straight paths

Answer

Maximize a classifier's log-likelihood

Answer

Add discrete noise tokens

Answer

Train a contrastive objective

Question 18

What is 'v-prediction' parameterization and why is it used?

Accepted Answer

v = sqrt(alpha_bar_t)*epsilon - sqrt(1-alpha_bar_t)*x_0, for stable targets at all SNRs

Answer

v = x_0 + epsilon

Answer

v = the classifier gradient

Answer

v = the posterior mean

Question 19

What is progressive distillation in diffusion models?

Accepted Answer

Student learns one step matching two teacher steps, halving steps repeatedly

Answer

Train a smaller network on the same number of steps

Answer

Distill the classifier into the U-Net

Answer

Remove the autoencoder

Question 20

What is the 'zero terminal SNR' issue and why does it matter for sampling?

Accepted Answer

Nonzero terminal SNR causes train/test mismatch limiting dynamic range; rescale to zero

Answer

It means the model overfits the prior

Answer

It only affects training speed

Answer

It is required for classifier guidance

Question 21

How is the reverse-process mean typically parameterized in DDPM given an epsilon network?

Accepted Answer

mu = (1/sqrt(1-beta_t))(x_t - (beta_t/sqrt(1-alpha_bar_t)) * eps_theta)

Answer

mu = x_t - eps_theta

Answer

mu = sqrt(alpha_bar_t) * eps_theta

Answer

mu = (x_t + eps_theta)/2

Question 22

What metric is most commonly used to evaluate image diffusion sample quality, and what does it measure?

Accepted Answer

FID: Frechet distance between Inception feature distributions of real and generated images

Answer

BLEU score on captions

Answer

Pixel-wise MSE to training images

Answer

Discriminator accuracy

Question 23

Why do diffusion models generally provide more stable training and better mode coverage than GANs?

Accepted Answer

A stable denoising regression avoids adversarial collapse, improving mode coverage

Answer

They use a discriminator that prevents collapse

Answer

They sample in one step like GANs

Answer

They do not require any neural network

Question 24

What is SDEdit (image-to-image) and how does the noising strength parameter behave?

Accepted Answer

Noise to an intermediate step then denoise; higher strength = more change, less fidelity

Answer

Always start from pure noise ignoring the input

Answer

Strength controls only the guidance scale

Answer

Higher strength preserves the input more

Question 25

How does inpainting with a diffusion model typically enforce the known (unmasked) region?

Accepted Answer

Replace known region each step with noised original, let model fill the mask

Answer

Train a separate model per mask shape

Answer

Zero out the masked region and stop

Answer

Only condition via text, ignore the pixels

Question 26

What is the principle behind ControlNet for conditioning diffusion models on spatial inputs?

Accepted Answer

A trainable copy of the U-Net encoder with zero-init convs adds spatial control

Answer

It retrains the whole U-Net from scratch

Answer

It is a text encoder replacement

Answer

It is a new sampling scheduler

Question 27

What distinguishes a Diffusion Transformer (DiT) from a U-Net-based diffusion model?

Accepted Answer

A transformer backbone on latent patches with adaLN conditioning instead of a U-Net

Answer

A U-Net with extra skip connections

Answer

A GAN generator

Answer

An RNN denoiser

Question 28

How are diffusion models extended to text-to-video generation while keeping temporal coherence?

Accepted Answer

Add temporal (space-time) attention/conv layers over latent frame sequences

Answer

Generate each frame fully independently

Answer

Use a single 2D image and tile it

Answer

Rely only on optical flow, no diffusion

Question 29

Why does naively applying continuous Gaussian-diffusion math fail for discrete data like text, and what approaches address it?

Accepted Answer

Use discrete transition kernels (D3PM) or embedding-space diffusion instead of Gaussian noise

Answer

Add Gaussian noise directly to token IDs

Answer

Diffusion cannot model text at all

Answer

Use a GAN on tokens

Question 30

What are DPM-Solver / DEIS-type samplers and why are they faster than basic DDIM?

Accepted Answer

High-order exponential-integrator ODE solvers needing fewer steps for high quality

Answer

They add stochastic noise to speed convergence

Answer

They retrain the model with fewer timesteps

Answer

They are GAN-based refiners

Test yourself