Diffusion Models quiz

Question 1

What is the core idea behind diffusion models?

Accepted Answer

Learn to reverse a gradual noising process: add noise forward, denoise backward to generate

Answer

Train a generator against a discriminator

Answer

Encode data to latent then decode

Answer

Apply iterative random projections

Question 2

What is the denoising objective in DDPM?

Accepted Answer

MSE between predicted noise ε_θ(x_t,t) and actual added noise ε

Answer

Cross-entropy between x_0 and x_T

Answer

GAN adversarial loss

Answer

VAE ELBO only

Question 3

What is the noise schedule in diffusion models?

Accepted Answer

Cosine schedule: slower early noise addition, better FID vs linear schedule

Answer

Uniform schedule is always best

Answer

Noise schedule doesn't affect quality

Answer

Linear schedule adds too little noise

Question 4

What is the U-Net architecture in diffusion models?

Accepted Answer

Encoder-decoder with skip connections; time embedding injected via AdaGN

Answer

Pure MLP for denoising

Answer

No skip connections

Answer

Same architecture as ResNet

Question 5

What is score matching and its connection to diffusion?

Accepted Answer

Denoising score matching is equivalent to diffusion noise prediction: ε_θ = -√(1-ᾱ) s_θ

Answer

Score matching is unrelated to diffusion

Answer

Score function replaces the U-Net

Answer

Only for discrete data

Question 6

What is classifier guidance in diffusion models?

Accepted Answer

Modified score = unconditional score + guidance_scale × classifier gradient

Answer

Trains discriminator alongside diffusion model

Answer

Only works for text conditioning

Answer

Classifier gradient is added to the loss

Question 7

What is DALL-E 2 and how does it work?

Accepted Answer

Text→CLIP text embedding→prior→CLIP image embedding→diffusion decoder→image

Answer

Direct text-to-pixel mapping

Answer

GAN-based generation

Answer

Uses VAE without diffusion

Question 8

What is CLIP and its role in text-to-image diffusion?

Accepted Answer

Jointly trains image+text encoders via contrastive loss; embeddings condition diffusion via cross-attention

Answer

Trains only an image encoder

Answer

Uses VAE not contrastive loss

Answer

Only for image classification

Question 9

What is DPM-Solver and why does it matter?

Accepted Answer

ODE solver for diffusion reverse process; enables 10-20 step high-quality sampling

Answer

A new noise schedule

Answer

A training objective

Answer

A type of U-Net architecture

Question 10

What is score-based generative model via SDEs?

Accepted Answer

Forward SDE adds noise; reverse SDE uses learned score to denoise; unifies all diffusion variants

Answer

SDE is only used for training

Answer

Replaces the U-Net

Answer

Only applies to continuous data

Question 11

What is ControlNet and what does it enable?

Accepted Answer

Trainable copy of U-Net encoder processes spatial control inputs added to frozen backbone

Answer

Replaces the full U-Net

Answer

Only for text conditioning

Answer

Fine-tunes all parameters

Question 12

What is the difference between DDPM and VQ-VAE for image generation?

Accepted Answer

DDPM: iterative noise-based; VQ-VAE: discrete codebook tokens + autoregressive; modern systems combine both in latent space

Answer

VQ-VAE and DDPM are identical

Answer

DDPM encodes to tokens

Answer

VQ-VAE uses Gaussian noise

Question 13

What is a Diffusion Transformer (DiT)?

Accepted Answer

Replaces U-Net with Vision Transformer; patches of latents processed by self-attention; scales better

Answer

Diffusion model with no architecture change

Answer

Uses RNN instead of attention

Answer

Only for class-conditional generation

Question 14

What is the variational lower bound (ELBO) in diffusion models?

Accepted Answer

ELBO decomposes into per-timestep KL terms between learned and true reverse posteriors

Answer

Same as standard VAE ELBO

Answer

MSE on x_0 directly

Answer

No variational objective needed

Question 15

What is noise prediction vs x_0 prediction vs v-prediction?

Accepted Answer

ε: predicts noise; x_0: predicts clean image; v: predicts linear combination; v is most stable

Answer

Only x_0 prediction is used

Answer

ε and x_0 prediction are identical

Answer

v-prediction replaces the U-Net

Question 16

What is consistency models?

Accepted Answer

Maps any x_t directly to x_0 in one step; trained via distillation or self-consistency

Answer

Needs T=1000 steps always

Answer

Approximate alternative to diffusion

Answer

Uses GAN adversarial training

Question 17

What are diffusion models used for beyond image generation?

Accepted Answer

Audio, video, 3D shapes, molecules, motion — any modality with appropriate denoising

Answer

Only 2D images

Answer

Only for generation, not restoration

Answer

Requires image-like 2D structure

Question 18

What is Stable Diffusion XL (SDXL) vs Stable Diffusion 1.x?

Accepted Answer

Larger UNet + dual CLIP encoders + 1024px native resolution + v-prediction vs SD1.x

Answer

Same architecture as SD1.x, just more steps

Answer

Uses GAN instead of diffusion

Answer

Only works with NVIDIA A100

Question 19

What is the role of text conditioning strength in Stable Diffusion?

Accepted Answer

Higher CFG → more prompt adherence; lower → more diversity; default ~7.5 balances both

Answer

Higher CFG always improves quality

Answer

CFG scale determines sampling steps

Answer

Only affects unconditional generation

Question 20

What is the inversion problem in diffusion models (DDIM inversion)?

Accepted Answer

Run DDIM forward numerically to get x_T, then edit prompt and denoise backward

Answer

Same as training from scratch

Answer

Requires VAE encoder

Answer

Only works for low-resolution images

Question 21

What is the signal-to-noise ratio (SNR) in diffusion models?

Accepted Answer

SNR = ᾱ_t/(1-ᾱ_t); Min-SNR weighting clips it to balance training across easy and hard timesteps

Answer

SNR is the learning rate ratio

Answer

SNR determines the number of diffusion steps

Answer

SNR = variance of noise

Question 22

What is the U-Net skip connection's role in diffusion denoising?

Accepted Answer

Preserve fine spatial details from encoder to decoder, critical for high-frequency denoising at small t

Answer

Only used in the generator

Answer

Replace attention in U-Net

Answer

Only active at t=0

Question 23

What is Dreambooth vs Textual Inversion for personalization?

Accepted Answer

Textual Inversion: learn new token embedding only; Dreambooth: fine-tune all weights — more expressive but costlier

Answer

Textual Inversion fine-tunes all weights

Answer

Dreambooth only changes text encoder

Answer

They require the same amount of training

Question 24

What is the forward process distribution q(x_t | x_0)?

Accepted Answer

q(x_t|x_0) = N(√ᾱ_t x_0, (1-ᾱ_t)I) — closed-form allows sampling any t directly

Answer

Must simulate all t steps sequentially

Answer

Requires numerical integration

Answer

Only defined for t=1

Question 25

What is masked diffusion for text generation?

Accepted Answer

Corrupts text by masking tokens progressively; trains denoiser to predict originals; enables parallel generation

Answer

Same as BERT masked LM

Answer

Uses Gaussian noise on token IDs

Answer

Cannot generate text

Question 26

What makes diffusion models better than GANs for image generation?

Accepted Answer

Stable training (MSE loss), no mode collapse, better mode coverage, easy conditioning — at cost of slower inference

Answer

GANs still produce higher quality

Answer

Diffusion requires discriminator

Answer

GANs cover modes better

Test yourself