Name: Harsh Santosh Gelda
PRN: 22UF17379CM017
Class: Btech4 16
Tutorial 6
Scenario-Based Question:
A startup company is developing an AI system to generate realistic human face images for a video game.
They have a small dataset of real faces and want the AI to generate new, unique, and realistic faces that
do not belong to real people.
During development, the AI team considers using Generative Adversarial Networks (GANs) or
Variational Autoencoders (VAEs).
As an AI engineer, answer the following:
1. Explain how GANs and VAEs generate new images.
2. Describe the main components and working process of both models.
3. Compare GANs and VAEs in terms of image quality, training stability, and applications.
4. Recommend which model the company should use for generating highly realistic game characters
and justify your choice.
Answer:
1. How GANs and VAEs Generate New Images
Both GANs and VAEs are generative models, meaning they learn the underlying distribution of training
data and generate new samples similar to it.
● GANs: Generate images through a competition between two neural networks.
● VAEs: Generate images by learning a compressed probabilistic representation (latent space) of
the data.
2. Components and Working Process
A. Generative Adversarial Networks (GANs)
GANs consist of two neural networks that compete with each other. Components
1. Generator (G)
○ Takes random noise (latent vector) as input.
○ Generates fake images that resemble real images.
2. Discriminator (D)
○ Acts as a binary classifier.
○ Determines whether an image is real (from dataset) or fake (from generator).
Working Process
1. Generator creates fake images from random noise.
2. Discriminator receives both real and generated images.
3. Discriminator tries to correctly classify them.
4. Generator learns to fool the discriminator by producing more realistic images.
5. The process continues until generated images become highly realistic.
Key Idea
The system works like a game between a counterfeiter and a police officer.
● Generator → Counterfeiter
● Discriminator → Police
B. Variational Autoencoders (VAEs)
VAEs are probabilistic generative models based on autoencoders.
Components
1. Encoder
○ Compresses input data into a latent distribution (mean and variance).
2. Latent Space
○ A continuous probability space representing learned features.
3. Decoder
○ Reconstructs data from the latent vector to produce new samples.
Working Process
1. Input image passes through the encoder.
2. Encoder maps image into latent distribution (μ, σ).
3. A sample is drawn from this distribution.
4. The decoder reconstructs the image.
5. Training minimizes:
○ Reconstruction loss
○ KL divergence (regularization).
3. Comparison of GANs and VAEs
Feature GANs VAEs
Architecture Generator + Discriminator Encoder + Decoder
Output Quality Very sharp and realistic images Often blurry images
Training Stability Hard to train More stable
Latent Space Not explicitly structured Well-organized latent space
Applications Image generation, deepfakes, art Image reconstruction, anomaly detection
creation
4. Recommended Model for the Company
For generating highly realistic human faces, the company should use GANs. Reasons
1. Higher image quality
○ GANs generate sharp and photorealistic images.
2. Widely used in image synthesis
○ Models like StyleGAN and DCGAN produce extremely realistic faces.
3. Better for visual realism
○ VAEs tend to generate blurrier images.
Final Recommendation
The company should implement GANs, especially advanced variants like StyleGAN, to generate high-
quality and realistic video game character faces.