Case study terminology
Every term from the booklet's "Additional terminology" list, explained in exam-ready words, plus a few extra terms that help you write about the case study. Use them precisely in your answers.
Standard level and higher level
Steps that reduce unfair or unrepresentative patterns in a model's training data and outputs: auditing and rebalancing the dataset, testing outputs across many prompts, and human review before publishing.
Generating a recurring character or motif with the same appearance across many images, e.g. a brand mascot in every advert of a campaign.
A neural network whose layers slide small filters over an image to detect features (edges, then shapes, then objects). In many diffusion models the denoiser is a CNN (a U-Net); some newer ones use transformers.
Deliberately choosing, checking, cleaning and documenting training data, for example to exclude copyrighted or personal images and to make the data representative.
Removing noise from an image. In a diffusion model a neural network predicts the noise in an image so that part of it can be subtracted, step by step.
The method that modern diffusion models are based on: in training, noise is gradually added according to a fixed schedule; a network learns to gradually remove it, which is then used to generate images from noise.
A generative model that creates an image by starting from random noise and refining it over many denoising steps; it produces photorealistic images but needs a lot of computation.
Representing something (a word, a style, a character) as a vector of numbers and giving that same vector to the model each time, e.g. to keep a character consistent across images.
Producing images guided by a specific input (the condition), such as a text prompt, a class label, a sketch or a segmentation map, so the output matches an objective or style.
Producing images without any input except random noise, relying only on patterns learned from the training data; useful for exploration and synthetic datasets.
Turning one image into another form while keeping its structure, e.g. a sketch into a realistic rendering, or a black-and-white photo into colour.
Adding random (Gaussian) noise: to training images in the forward process, and as the random starting point from which a diffusion model generates a new image.
An image in which each region is labelled (usually by colour) with what it contains, such as sky, road or person; a conditional model can turn it into a realistic image with that layout.
Generating images from a natural-language prompt, e.g. DALL·E or Stable Diffusion; a text encoder turns the prompt into an embedding that guides generation.
How reliably training settles on a good model. Unstable training may oscillate, fail to improve or produce repetitive or poor outputs.
Higher level only HL
HL students should know diffusion models, GANs and hybrid models in depth, and VAEs and flow-based models broadly.
The competition between a GAN's generator and discriminator: the generator improves by learning to deceive the discriminator, and the discriminator improves at telling real from fake.
A list of D random numbers given to a GAN's generator as input; each different vector produces a different image.
A generative model made of two neural networks trained together in competition: a generator that makes synthetic images and a discriminator that judges whether images are real or fake.
The network in a GAN that classifies an image as real (from the training dataset) or fake (from the generator).
The network in a GAN that transforms a noise vector into a candidate image, and is trained to fool the discriminator.
A model that combines different types of generative AI (e.g. a VAE with a diffusion model) to overcome the limitations of a single approach: higher quality, more control, more flexibility.
A generative model that learns an exact, invertible mapping between random noise and images, so generation can be reversed without loss and each image traced; useful when precision, reproducibility or transparency matter.
A model with an encoder that compresses images into a latent space and a decoder that turns points in that space into images; good for flexible exploration but often blurry.
A compressed, organised representation of the data in which similar images are placed close together; moving through it changes the generated image smoothly.
A GAN failure in which the generator produces only a limited variety of outputs instead of capturing the full diversity of the dataset.
Useful extra terms (not in the booklet)
AI that creates new content (images, text, audio) resembling its training data, rather than only classifying or predicting.
A vector of numbers that represents the meaning of something; similar meanings give similar vectors.
The text instruction given to a text-to-image model.
The number that sets the starting random noise. The same seed, prompt and settings reproduce the same image.
A CNN architecture widely used as the denoiser in diffusion models (e.g. Stable Diffusion 1 and 2): it shrinks the image to capture overall structure, then expands it back to full detail.
Running the diffusion process on a compressed latent version of the image (made by a VAE) to save computation; used by Stable Diffusion.
A setting that controls how strongly generation follows the prompt: higher means closer to the prompt but less variety.
Further training of a pre-trained model on a smaller, specific dataset (e.g. a studio's own artwork); LoRA is a cheap way to do this.
Training a smaller or faster 'student' model to copy a slower 'teacher', e.g. a diffusion model that needs only 1–4 steps.
Artificially generated data used to train other AI models.
Realistic AI-generated or altered media showing real people saying or doing things they never did.
A standard for attaching signed information to a file about how it was created and edited, including AI use.
Practise: flashcards and quick quiz
These use the booklet's terms only (HL terms are left out when HL is hidden).