05.1 The generator, discriminator and noise vector (HL only)

01 | TWO NETWORKS WITH DIFFERENT JOBS

A generative adversarial network , or GAN, learns through an interaction between two networks. The generator produces synthetic candidates. The discriminator learns to distinguish real training examples from generated examples.

The 2027 case study includes GANs as HL-only content. Visionary Studios is considering them as one possible image generation approach.

Start by keeping the roles separate: the generator creates a candidate, while the discriminator assesses inputs for its real-versus-generated learning task. Neither is a human designer.

02 | WHAT DOES THE GENERATOR RECEIVE?

In a basic unconditional GAN, the generator takes a sampled numerical vector, usually written z . The vector supplies a starting input that can lead to different outputs.

The generator applies learned calculations to transform that vector into an image-like output. Its learned parameters determine how input values influence the result.

This introductory setup does not require a text prompt. Conditional GANs can also use guidance such as labels, but that is an extension of the basic arrangement.

03 | WHAT IS A VECTOR?

A vector is an ordered list of numerical components. For example, [0.4, −1.2, 0.7, 0.1] contains four components.

A D-dimensional noise vector has D components. D does not mean the number of image pixels, objects or output channels. A generator can transform a relatively short vector into a much larger image array.

The booklet’s term describes the size of the generator’s starting numerical input, not a physical direction in the classroom.

04 | WHERE DO THE NOISE VALUES COME FROM?

Components are sampled from a chosen distribution, such as a standard normal distribution or a specified uniform distribution. The particular choice belongs to the model setup.

A pseudorandom process can use a seed for repeatable sampling under compatible conditions. Different sampled vectors can produce different candidates.

Random input alone cannot create meaningful images. The trained generator supplies learned relationships that shape the output.

05 | THE NOISE VECTOR IS NOT A SMALL PHOTOGRAPH

The entries of z need not correspond directly to image locations or human-readable attributes. The first number is not automatically hair colour, and the second is not automatically pose.

The generator learns a mapping from input space to output space. Some learned directions may influence meaningful properties, but such interpretation is not guaranteed for every coordinate.

The vector does not contain a particular photograph hidden beneath noise. It is a numerical input used to sample from learned generative behaviour.

06 | FROM VECTOR TO IMAGE

The generator transforms its input through layers of learned operations. Depending on the architecture, these can build up spatial representations until the network produces an image.

For a 64 × 64 RGB output, the image has 64 × 64 × 3 = 12,288 channel values . The input could, for example, have 100 components. Those are illustrative dimensions, not values specified for Visionary Studios.

The network’s parameters and operations perform the transformation. Merely copying each input component repeatedly would not explain learned image generation.

07 | WHAT DOES THE DISCRIMINATOR RECEIVE?

In the basic GAN, the discriminator receives an image candidate. That candidate can come from the real training dataset or from the generator.

The discriminator is trained using the source label: the training process knows whether an example was drawn from the dataset or produced by the generator.

It does not normally receive the original noise vector as an answer key. Its task concerns the image input and the chosen real-versus-generated objective.

08 | WHAT DOES ITS OUTPUT MEAN?

In the original probability-style GAN formulation, the discriminator produces a value interpreted as an estimate of how likely the input is to belong to the real-data class.

A high real-class score means the discriminator currently treats that input as real-like under its learned task. It is not a score for copyright compliance, client approval or factual truth.

GAN variants can use different outputs and losses. This lesson uses the standard introductory probability-style explanation rather than treating it as universal.

09 | REAL DATA DOES NOT MEAN VERIFIED TRUTH

“Real” is a training label for examples drawn from the selected dataset. It does not guarantee that every example is an unedited photograph or a factually accurate scene.

A dataset of illustrations can provide the real-data class for an illustration GAN. The contrast is dataset examples versus generator-produced examples.

Dataset selection therefore matters. Visionary Studios must evaluate what the learning examples represent rather than assume “real” means ethically cleared or suitable for every campaign.

10 | CONNECT THE TWO PATHS

The discriminator sees both types of image during training. It does not send a natural-language description of defects back to the generator.

11 | HOW THE DISCRIMINATOR LEARNS

The discriminator predicts on real and generated examples. A loss compares its predictions with the known source labels.

Backpropagation computes gradients, and an optimizer updates discriminator parameters. Generated images are treated as fixed examples for this discriminator update; the generator is not updated by that step.

The objective is to distinguish the sources more effectively. It is not to reward every image that a person finds attractive.

12 | HOW THE GENERATOR RECEIVES A LEARNING SIGNAL

For a generator update, its image is passed through the discriminator. A generator objective encourages outputs that receive a more favourable real-like assessment.

Gradients can be computed through the discriminator’s operations back into the generator. During that update, discriminator parameters are held fixed while generator parameters are adjusted.

Holding the discriminator fixed does not mean its computation disappears. It still provides a differentiable route for the learning signal. The next lesson develops the alternating process.

13 | THEY DO NOT BOTH UPDATE IN THE SAME WAY

The two networks have different objectives and parameter sets. Training commonly alternates update phases rather than treating them as one network with one identical target.

The generator tries to improve its synthetic outputs under the adversarial objective. The discriminator tries to improve its distinction between dataset and generated images.

The balance can be difficult. A useful learning signal is not guaranteed simply because two networks are connected; later lessons examine instability and mode collapse.

14 | A SCORE CAN BE MISLEADING

Imagine a generated image receiving a real-class score of 0.9. It has fooled this discriminator strongly, but might still contain a wrong logo or fail an advertising layout.

A weak discriminator can also assign favourable scores to poor outputs. Conversely, a high classification accuracy for the discriminator alone does not prove that the whole GAN is progressing usefully.

Inspect generated quality and diversity independently. A discriminator is a training component, not the final creative or ethical approval authority.

15 | WHAT IS NEEDED AFTER TRAINING?

In a basic GAN deployment, the trained generator can produce images from sampled vectors without running the discriminator for each generation request.

The discriminator was used to supply learning pressure during training. A deployment may use additional review tools, but that is a separate workflow decision.

For a conventional feed-forward GAN, a candidate can be produced in one generator forward pass. This differs from the repeated denoising process studied earlier, although actual system latency still depends on the model and supporting work.

16 | CHANGING THE VECTOR AND CHANGING THE MODEL

With fixed generator parameters, a new vector can produce a new sample. That is generation, not retraining.

With the same vector, changing the learned parameters can change the output. Recording a vector or seed therefore does not establish repeatability across different checkpoints or software conditions.

Also distinguish variety from usefulness: many different vectors do not guarantee that a poorly trained generator produces a diverse set of acceptable images.

17 | APPLY IT TO VISIONARY STUDIOS

The booklet does not identify a selected GAN architecture or noise-vector size. Use numerical examples as illustrations rather than company facts.

18 | TRACE THE ROLES

Draw the real-data path and the generated path. Label the noise vector, generator, generated image, discriminator and source labels. Then explain which parameters change during each update phase.

Finish by describing what the basic generator needs to create a new image after training.

HL only. Read alongside the GAN section of the official 2027 case study. The dimensions and scores in this lesson are independent teaching examples.

CHECK YOUR UNDERSTANDING | TRACE THE GAN COMPONENTS