Advanced Diffusion Models quiz

Question 1

What is the mathematical objective of denoising diffusion probabilistic models (DDPM)?

Accepted Answer

DDPM trains a network to predict the noise added at each timestep via MSE loss derived from the ELBO

Answer

DDPM uses GAN discriminator loss to distinguish real from generated images

Answer

DDPM minimizes KL divergence between the generated and real image distributions directly

Answer

DDPM trains by reconstructing the clean image from partial observations

Question 2

What is classifier-free guidance (CFG) and how does it trade off diversity and quality?

Accepted Answer

CFG scales the difference between conditional and unconditional scores; higher guidance improves adherence but reduces diversity

Answer

CFG requires a separately trained image classifier at inference time

Answer

CFG only works with text conditioning, not image conditioning

Answer

CFG was introduced in Stable Diffusion XL, not the original DDPM

Question 3

What is the flow matching framework and how does it relate to diffusion models?

Accepted Answer

Flow matching uses straight-line optimal transport paths between noise and data, enabling fewer sampling steps than standard diffusion

Answer

Flow matching is identical to DDPM but with a different loss function name

Answer

Flow matching requires a separate discriminator network during training

Answer

Flow matching cannot be used for text-to-image generation

Question 4

How does the DiT (Diffusion Transformer) architecture replace the traditional U-Net in diffusion models?

Accepted Answer

DiT replaces the U-Net with a Vision Transformer using adaptive layer norm conditioning on timestep and class embeddings

Answer

DiT uses convolutional layers for spatial processing and transformers only for text encoding

Answer

DiT requires a separate encoder for images and a separate decoder for conditioning

Answer

DiT was introduced as an extension of U-Net with attention at each resolution

Question 5

What is Consistency Models and how do they achieve single-step generation?

Accepted Answer

Consistency Models train a function to map any noisy sample on a trajectory to the clean origin, enabling single or few-step generation

Answer

Consistency Models require 1000 steps like DDPM but use a different loss

Answer

Consistency Models are only applicable to text-to-image, not video

Answer

Consistency Models replace the U-Net with a GAN discriminator

Question 6

What is negative prompting in diffusion models and how does it interact with CFG mathematically?

Accepted Answer

Negative prompting replaces the unconditional CFG score with a negative-prompt-conditioned score, steering away from unwanted features

Answer

Negative prompting requires training a separate classifier to reject bad images

Answer

Negative prompting is implemented as post-processing filter after image generation

Answer

Negative prompts reduce guidance scale automatically

Question 7

What evaluation metrics are standard for text-to-image generation quality?

Accepted Answer

FID measures distributional quality; CLIP Score measures text-image alignment; both are standard T2I evaluation metrics

Answer

BLEU and ROUGE are the primary metrics for text-to-image evaluation

Answer

Only human evaluation is valid for assessing diffusion model quality

Answer

IS (Inception Score) is the current gold standard metric

Question 8

What is the role of the VAE in Stable Diffusion and what happens if the VAE decoder has artifacts?

Accepted Answer

The VAE compresses images to a latent space for diffusion processing; VAE quality directly impacts pixel-level artifact presence

Answer

The VAE is only used during training, not at inference time

Answer

The VAE replaces the U-Net in Stable Diffusion

Answer

Stable Diffusion doesn't use a VAE—it operates directly on pixels

Question 9

How is MultiDiffusion used for panorama and high-resolution image generation without retraining?

Accepted Answer

MultiDiffusion tiles overlapping windows, independently denoises each, and averages predictions in overlaps for arbitrary resolution generation

Answer

MultiDiffusion trains multiple U-Nets, each specialized for a different resolution

Answer

MultiDiffusion requires video data to generate panoramic images

Answer

MultiDiffusion only works with DDIM, not DDPM samplers

Question 10

How does Sora (OpenAI) represent and generate video as spacetime patches?

Accepted Answer

Sora compresses videos to spacetime patches and processes them with a DiT transformer with full spacetime attention

Answer

Sora generates video frame-by-frame using a separate image diffusion model for each frame

Answer

Sora uses a recurrent LSTM to maintain temporal coherence between frames

Answer

Sora requires separate depth maps as input to maintain 3D consistency

Question 11

What is the key difference between deterministic and stochastic DDIM sampling?

Accepted Answer

Deterministic DDIM (eta=0) produces reproducible results enabling latent interpolation; stochastic DDIM (eta>0) adds noise for diversity

Answer

Stochastic DDIM uses more sampling steps than deterministic DDIM

Answer

Deterministic DDIM requires a separate classifier for conditioning

Answer

Both DDIM variants produce identical outputs regardless of eta

Test yourself