NEW
Font size
WorksheetsGenerative Models for Image & Video Recognition and Analysis
Total questions: 25
Worksheet time: 13mins
Generative models primarily learn which distribution?
Conditional distribution p(y∣x)
Joint distribution p(x,y) or data distribution p(x)
Prior distribution only
Deterministic mapping
Which generative model involves a generator and a discriminator competing?
VAE
Diffusion Model
GAN
Autoregressive Model
The role of the GAN discriminator is to:
Generate synthetic samples
Add noise to real data
Identify real vs fake samples
Remove noise
StyleGAN, CycleGAN, and VideoGAN are examples of:
Flow-based models
GAN variants
Diffusion models
Autoregressive models
VAEs learn a latent representation using:
One-stage CNNs
Encoder–decoder architecture
Autoregressive tokenization
Residual networks
In a VAE, the bottleneck contains:
Output images
Labels
Latent space (compressed embedding)
Reconstruction loss
Which VAE variant uses discrete codebook vectors?
β -VAE
CVAE
VQ-VAE
VAE-GAN
Autoregressive models generate data:
In a single forward pass
Randomly from noise
Pixel-by-pixel or frame-by-frame
Using invertible functions
PixelCNN is a classic architecture for:
Flow-based modeling
Autoregressive image generation
Diffusion-based denoising
GAN classification
Video autoregressive models typically add:
Color channels
Temporal sequence modeling
Noise scheduling
Mode collapse prevention
Diffusion models generate data by:
Deterministic reconstruction
Competing networks
Gradually denoising random noise
Sampling discrete tokens
The forward diffusion process:
Removes noise from images
Adds Gaussian noise over many steps
Classifies images
Generates high-resolution images
DDPMs contain:
Only a generator network
A forward and reverse Markov chain
A classifier and encoder
An invertible flow
Score-based generative models estimate:
Latent priors
Score functions of noisy distributions
Class probabilities
Flow Jacobians
Flow-based models rely on:
Invertible transformations
Adversarial training
Sequential denoising
Label conditioning
RealNVP is known for introducing:
Additive coupling
Autoregressive masking
Affine coupling layers
Temporal modeling
Glow adds which architectural component?
Invertible 1×1 convolution
GAN discriminator
Score matching
Recurrent modules
Continuous Normalizing Flows (CNFs) use:
Skip connections
Ordinary differential equations (ODEs)
Pixel-level dependencies
KL divergence
Flow-based models enable:
Lossless reconstruction
Only one-way transformations
Label-only predictions
Deterministic sampling only
A key application of flow models in video is:
Text summarization
Frame prediction
Speech synthesis
Optical character recognition
Hybrid generative–discriminative models combine:
Only discriminative learning
GAN with RNN
Generative + discriminative components
Reinforcement + supervised models
AC-GAN stands for:
Auto-Coupling GAN
Auxiliary Classifier GAN
Adaptive Convolutional GAN
Attention-Control GAN
Hybrid VAE–GAN models combine:
Pure diffusion + classifier
VAE stability + GAN realism
RNN + CNN
PixelCNN + VQ encoders
Generative models support recognition by:
Reducing image resolution
Domain adaptation and data augmentation
Removing training labels
Bypassing neural networks entirely
Diffusion models in recognition tasks help with:
Improving robustness via denoising
Generating synthetic labeled data
Reducing parameter count
Removing temporal dependencies
