wayground logo

Free Printable Worksheets

NEW

Font size

S
M
L
XL
Worksheets

Generative Models for Image & Video Recognition and Analysis

Total questions: 25

Worksheet time: 13mins

Name
Class
Date
1.

Generative models primarily learn which distribution?

a)

Conditional distribution p(yx)p(y\mid x)

b)

Joint distribution p(x,y)p(x,y) or data distribution p(x)p(x)

c)

Prior distribution only

d)

Deterministic mapping

2.

Which generative model involves a generator and a discriminator competing?

a)

VAE

b)

Diffusion Model

c)

GAN

d)

Autoregressive Model

3.

The role of the GAN discriminator is to:

a)

Generate synthetic samples

b)

Add noise to real data

c)

Identify real vs fake samples

d)

Remove noise

4.

StyleGAN, CycleGAN, and VideoGAN are examples of:

a)

Flow-based models

b)

GAN variants

c)

Diffusion models

d)

Autoregressive models

5.

VAEs learn a latent representation using:

a)

One-stage CNNs

b)

Encoder–decoder architecture

c)

Autoregressive tokenization

d)

Residual networks

6.

In a VAE, the bottleneck contains:

a)

Output images

b)

Labels

c)

Latent space (compressed embedding)

d)

Reconstruction loss

7.

Which VAE variant uses discrete codebook vectors?

a)

β\beta -VAE

b)

CVAE

c)

VQ-VAE

d)

VAE-GAN

8.

Autoregressive models generate data:

a)

In a single forward pass

b)

Randomly from noise

c)

Pixel-by-pixel or frame-by-frame

d)

Using invertible functions

9.

PixelCNN is a classic architecture for:

a)

Flow-based modeling

b)

Autoregressive image generation

c)

Diffusion-based denoising

d)

GAN classification

10.

Video autoregressive models typically add:

a)

Color channels

b)

Temporal sequence modeling

c)

Noise scheduling

d)

Mode collapse prevention

11.

Diffusion models generate data by:

a)

Deterministic reconstruction

b)

Competing networks

c)

Gradually denoising random noise

d)

Sampling discrete tokens

12.

The forward diffusion process:

a)

Removes noise from images

b)

Adds Gaussian noise over many steps

c)

Classifies images

d)

Generates high-resolution images

13.

DDPMs contain:

a)

Only a generator network

b)

A forward and reverse Markov chain

c)

A classifier and encoder

d)

An invertible flow

14.

Score-based generative models estimate:

a)

Latent priors

b)

Score functions of noisy distributions

c)

Class probabilities

d)

Flow Jacobians

15.

Flow-based models rely on:

a)

Invertible transformations

b)

Adversarial training

c)

Sequential denoising

d)

Label conditioning

16.

RealNVP is known for introducing:

a)

Additive coupling

b)

Autoregressive masking

c)

Affine coupling layers

d)

Temporal modeling

17.

Glow adds which architectural component?

a)

Invertible 1×11\times 1 convolution

b)

GAN discriminator

c)

Score matching

d)

Recurrent modules

18.

Continuous Normalizing Flows (CNFs) use:

a)

Skip connections

b)

Ordinary differential equations (ODEs)

c)

Pixel-level dependencies

d)

KL divergence

19.

Flow-based models enable:

a)

Lossless reconstruction

b)

Only one-way transformations

c)

Label-only predictions

d)

Deterministic sampling only

20.

A key application of flow models in video is:

a)

Text summarization

b)

Frame prediction

c)

Speech synthesis

d)

Optical character recognition

21.

Hybrid generative–discriminative models combine:

a)

Only discriminative learning

b)

GAN with RNN

c)

Generative + discriminative components

d)

Reinforcement + supervised models

22.

AC-GAN stands for:

a)

Auto-Coupling GAN

b)

Auxiliary Classifier GAN

c)

Adaptive Convolutional GAN

d)

Attention-Control GAN

23.

Hybrid VAE–GAN models combine:

a)

Pure diffusion + classifier

b)

VAE stability + GAN realism

c)

RNN + CNN

d)

PixelCNN + VQ encoders

24.

Generative models support recognition by:

a)

Reducing image resolution

b)

Domain adaptation and data augmentation

c)

Removing training labels

d)

Bypassing neural networks entirely

25.

Diffusion models in recognition tasks help with:

a)

Improving robustness via denoising

b)

Generating synthetic labeled data

c)

Reducing parameter count

d)

Removing temporal dependencies