IB CS HL case study terms 2027

diffusion model

A generative model that creates an image by beginning with noise and repeatedly removing it. It is central because the first shared challenge concerns managing iterative denoising and its computational demands.

denoising

The repeated process of predicting and removing noise so that meaningful visual features emerge. Explain it as a sequence of refinements, not a single cleaning operation.

denoising diffusion probabilistic model (DDPM)

A probabilistic framework involving a forward process that adds noise to training images and a learned reverse process that removes noise during generation. It gives technical precision to an explanation of diffusion.

noise injection

The deliberate addition or use of random noise within the diffusion process. Random starting noise allows the model to produce different outputs rather than reproduce one fixed image.

dataset curation

Selecting, checking, cleaning, documenting, and managing training data. This directly affects copyright risk, representation, model quality, and the second shared challenge.

bias mitigation

Measures intended to identify and reduce systematically unfair or distorted outputs. Examples include improving dataset balance, testing outputs across demographic groups, and applying human review.

convolutional neural network (CNN)

A neural network suited to image data because learned filters detect spatial features such as edges, textures, and shapes. The case study presents a CNN as the typical neural denoiser.

training stability

The extent to which training progresses reliably rather than oscillating, diverging, or producing unusable results. It is particularly useful when comparing diffusion models with GANs.

image generation

The creation of new visual data using a generative model trained to learn patterns from existing data. This is the case study's broad task, but answers need more specific terminology to gain depth.

text-to-image generation

Producing an image conditioned on a written prompt. For Visionary Studios, it could turn a client's written advertising brief into concept images.

conditional

Generation guided by additional information such as text, a class label, an image, or a segmentation map. Conditioning improves control and alignment with a design brief.

character consistency

Maintaining recognisable identity, clothing, proportions, or style for the same character across multiple images. This matters for mascots, storyboards, and campaign continuity.

embedding-based approach

Representing a concept, style, or identity as a numerical vector that a model can use during generation. Such an approach may help preserve a recurring character's features.

image-to-image translation

Transforming an input image into a related output while preserving selected structural features. A rough concept sketch could become a polished advertising image.

segmentation map

A labelled spatial representation identifying regions such as person, sky, road, or product. It can condition generation so that objects appear in specified locations.

unconditional

Generation without a prompt, label, or other explicit condition, usually from random noise alone. It is useful for open-ended exploration but offers less control over a commercial brief.

generative adversarial network (GAN)

An architecture in which a generator and discriminator train competitively. It is the basis of the third challenge and a major comparison with diffusion models.

(Do not describe a GAN merely as "two AIs competing." State what each network does and explain the feedback relationship.)

generator

The GAN network that converts a random input vector into a synthetic image. Its objective is to create outputs the discriminator classifies as real.

discriminator

The GAN network trained to distinguish genuine training samples from generated ones. Its feedback drives improvement in the generator.

hybrid model

A system combining two or more architectures to exploit complementary strengths. The fourth challenge asks HL students to evaluate combinations involving VAEs, GANs, flow-based models, and diffusion models.

(A hybrid is not automatically superior: additional components can increase computational cost, implementation difficulty, and training complexity.)

adversarial dynamic

The competitive training relationship in which the generator attempts to deceive the discriminator while the discriminator improves its detection. Poor balance can destabilise training.

mode collapse

A GAN failure in which the generator produces only a narrow range of outputs despite diverse possible inputs. For a design company, this limits creativity and campaign variety.

latent space

A lower-dimensional learned representation in which meaningful visual properties are encoded as numerical positions or directions. Manipulating it can support controlled variation and interpolation.

variational autoencoder (VAE)

A generative architecture that learns a structured probability distribution in latent space, then samples from it to create data. VAEs support smooth interpolation but may produce less sharp images than some alternatives.

flow-based model

A generative model using reversible transformations between a simple probability distribution and complex data. It can provide exact likelihood calculations but may require substantial memory and architectural complexity.

D-dimensional noise vector

A random vector containing D numerical components that provides the generator's starting input. Different vectors can lead to different generated images.