Introduction to Conditional GANs (cGANs)

Standard Generative Adversarial Networks, as reviewed in Chapter 1, learn to map a random noise vector $z$ to a sample $x$ that resembles data from a target distribution $p_{data}(x)$. While powerful, this process offers little explicit control over the type of sample generated. If you train a GAN on images of handwritten digits (like MNIST), it will generate realistic digits, but you cannot simply ask it to produce, say, a '7'. You get whichever digit the sampled $z$ happens to map to.

Conditional GANs (cGANs) extend the basic GAN framework to address this limitation by incorporating auxiliary information, often called a condition or label, denoted by $y$. This condition $y$ provides context or guidance to the generation process, allowing us to direct the output. Instead of learning the overall data distribution $p(x)$, a cGAN learns the conditional distribution $p(x|y)$.

The Core Idea: Conditioning the Generation

The fundamental change in a cGAN is that both the generator and the discriminator receive the conditional information $y$ as an additional input.

• Generator: The generator's task is now to produce a sample $x_{fake}$ that is realistic and matches the given condition $y$. Its input becomes not just the noise vector $z$, but the pair $(z, y)$. We denote the generator function as $G(z, y)$.
• Discriminator: The discriminator must now distinguish real data samples $x_{real}$ from generated samples $x_{fake}$, but it must do so in the context of the condition $y$. It needs to determine if the input pair $(x, y)$ is a valid pair, meaning $x$ is realistic and corresponds to the label $y$. Real pairs $(x_{real}, y)$ come from the true data distribution, while fake pairs $(G(z, y), y)$ come from the generator. We denote the discriminator function as $D(x, y)$.

This conditioning mechanism allows for targeted generation. If $y$ represents a digit class in MNIST, providing $y=7$ to $G(z, y)$ instructs the generator to specifically synthesize an image of a '7'.

Data flow in a standard GAN versus a Conditional GAN (cGAN). The cGAN incorporates conditional information $y$ into both the generator and discriminator.

Modifying the Objective Function

The introduction of the condition $y$ requires a modification to the standard GAN minimax objective function. The objective now reflects the conditional nature of the generation and discrimination tasks. The value function $V(D, G)$ for a cGAN becomes:

$$\min_G \max_D V(D, G) = \mathbb{E}_{(x, y) \sim p_{data}(x, y)}[\log D(x, y)] + \mathbb{E}_{z \sim p_z(z), y \sim p_y(y)}[\log(1 - D(G(z, y), y))]$$

Let's break this down:

1. $\mathbb{E}_{(x, y) \sim p_{data}(x, y)}[\log D(x, y)]$: The discriminator $D$ aims to maximize this term. It tries to assign a high probability (close to 1) to real data samples $x$ when paired with their correct corresponding condition $y$. The expectation is taken over real data samples $x$ and their associated conditions $y$ drawn from the joint data distribution $p_{data}(x, y)$.
2. $\mathbb{E}_{z \sim p_z(z), y \sim p_y(y)}[\log(1 - D(G(z, y), y))]$: The generator $G$ aims to minimize this term (while $D$ tries to maximize it by making $D(G(z,y), y)$ small). $G$ takes random noise $z$ and a condition $y$ (sampled from its distribution $p_y(y)$, which is often derived from the training data labels) to produce a fake sample $G(z, y)$. The discriminator $D$ then evaluates this generated sample along with its intended condition $y$. $D$ tries to assign a low probability (close to 0) to these fake pairs, while $G$ tries to fool $D$ into assigning a high probability.

Essentially, the minimax game remains, but it's now played over conditional distributions. The discriminator learns to model $p(x|y)$, differentiating between real conditional samples and fake ones. The generator learns to produce samples that are indistinguishable from real samples given the condition $y$.

How is Conditioning Information Incorporated?

The specific mechanism for feeding $y$ into $G$ and $D$ depends on the nature of $y$ and the network architectures. Common approaches, which we will explore further in the "Architectures for cGANs" section, include:

• Concatenation: If $y$ is a simple vector (e.g., a one-hot encoded class label or an embedding), it can be directly concatenated to the noise vector $z$ for the generator, or concatenated to the input $x$ (or intermediate feature maps) for the discriminator.
• Embedding Layers: Categorical labels $y$ are often passed through embedding layers to obtain dense vector representations before being combined with $z$ or $x$.
• Modulation Techniques: More advanced architectures might use $y$ to modulate weights or activations within the networks (e.g., Conditional Batch Normalization).

Why Use cGANs?

Conditional GANs provide a powerful handle for controlling generative processes. Instead of sampling randomly from the learned distribution, you can explicitly request specific types of outputs. This is immensely useful in various applications:

• Class-Conditional Image Synthesis: Generating images belonging to specific categories (e.g., generating a 'Siamese cat' image instead of just any cat).
• Text-to-Image Synthesis: Generating images based on textual descriptions (e.g., StackGAN, covered later in this chapter). Here, $y$ is a vector representation of the input text.
• Image-to-Image Translation: Translating an image from one domain to another based on a target domain label or example (e.g., converting sketches to photos).
• Attribute Manipulation: Modifying specific attributes of an image (e.g., adding glasses to a face image).

By conditioning the generation, cGANs move from simple mimicry towards controllable and directed synthesis, opening up a wider range of practical applications for generative modeling. The next sections will detail specific architectures and related techniques like InfoGAN, which seeks to learn meaningful conditioning factors automatically.