Skip to content
ExamSpine

Case study glossary

65 terms you should be able to explain without notes, in plain English. Each links to the topic that explains it in full.

A

Attribution
Crediting the people whose work was used. Hard for generators, because each image draws on many training images at once. Learn more

B

Backpropagation
The method that works out how much each weight in a network contributed to the error, so gradient descent can adjust it. Learn more
Bias (in AI output)
A systematic unfairness or skew in what a model produces, usually learned from unbalanced training data. Learn more
Bias audit
A structured test of a model's outputs, for example generating many images for neutral prompts and counting who and what appears. Learn more

C

Checkpoint
A saved copy of a model's weights during training, so the best version can be kept if training later gets worse. Learn more
Classifier-free guidance
A technique that compares the denoiser's prediction with and without the prompt and pushes further towards the prompt. Its strength is the guidance scale. Learn more
CLIP score
A measure of how well an image matches a text prompt, using a model trained on image and caption pairs. Higher is better. Learn more
Conditional GAN
A GAN where both networks also receive extra information, such as a label or text embedding, so you can ask for a particular kind of image. Learn more
Convolutional neural network (CNN)
A neural network that processes images by sliding small learned kernels across them, sharing weights across the whole image. Learn more
Cross-attention
Layers that let each part of the image being denoised look at the prompt's embedding, so the text steers the image. Learn more

D

Dataset curation
Choosing, filtering, balancing and documenting the data a model is trained on. Learn more
Deduplication
Removing repeated or near-identical images from a dataset, which reduces memorisation and over-weighting. Learn more
Diffusion model
A generative model trained to remove noise from images a step at a time. It creates new images by denoising pure random noise. Learn more
Discriminator
The GAN network that classifies images as real (from the training data) or fake (from the generator). Learn more
Distillation
Training a smaller or faster "student" model to copy a larger "teacher", for example a diffusion model that needs 1 to 4 steps instead of 50. Learn more
Downsampling
Shrinking feature maps (by pooling or a stride above 1) so later layers see a wider area of the image. Learn more

E

Embedding
A list of numbers that represents the meaning of something, such as a text prompt, so a network can use it. Learn more
Equilibrium (GAN)
The ideal end of GAN training, where generated images match the real distribution and the discriminator can only guess (0.5). Learn more

F

Fair use / copyright exceptions
Legal rules that allow some uses of copyrighted work without permission. Whether they cover AI training differs by country. Learn more
Feature map
The grid of values produced by sliding one kernel over an input; high where the kernel's pattern appears. Learn more
Feedback loop
When generated images are posted online and scraped into later training data, reinforcing the same defaults and biases. Learn more
FID (Fréchet Inception Distance)
A score comparing features of many real and many generated images. Lower means the generated set is closer to the real one. Learn more
Flow-based model
A generative model built from exactly invertible transformations, mapping noise to images and back, with exact likelihoods. Learn more
Forward process
The fixed process of adding noise to a training image over many steps until it becomes pure noise. Learn more

G

Gaussian noise
Random values drawn from a normal (bell-curve) distribution, added to images in the forward process. Learn more
Generative adversarial network (GAN)
A generator and a discriminator trained against each other: one makes images, the other tries to tell them from real ones. Learn more
Generative learning trilemma
The observation that GANs, VAEs and diffusion models each achieve only two of high quality, variety and fast generation. Learn more
Generator
The GAN network that turns a random latent vector into an image, trained to fool the discriminator. Learn more
GPU
A graphics processing unit: a processor with thousands of simple cores that does the parallel arithmetic of neural networks quickly. Learn more
Guidance scale
How strongly classifier-free guidance pushes an image towards the prompt. Higher follows the prompt more but gives less variety. Learn more

H

Hybrid model
A generative system that combines different kinds of model, such as an autoencoder with diffusion, or diffusion with a GAN. Learn more
Hyperparameter
A setting chosen before training rather than learned, such as the learning rate or batch size. Learn more

K

Kernel (filter)
A small grid of learned weights slid across an image to detect one feature, producing a feature map. Learn more

L

Latency
How long one request takes, from sending a prompt to getting the image. Learn more
Latent diffusion
Running diffusion on a compressed latent version of the image made by an autoencoder, then decoding it, so each step is much cheaper. Learn more
Latent space
A compressed space of numbers that represents images; nearby points give similar images. Learn more
Latent vector (z)
The random input to a GAN's generator. Different vectors give different images. Learn more
Licensed dataset
Training data used with the owners' permission, usually paid for, with a record of where each item came from. Learn more

M

Memorisation
When a model reproduces a training image almost exactly instead of making something new. Learn more
Minimax game
The GAN set-up where the discriminator tries to maximise how well it tells real from fake and the generator tries to minimise it. Learn more
Mode collapse
A GAN failure where the generator makes only a few kinds of output: high quality, low variety. Learn more
Model card / datasheet
A document recording how a model or dataset was made, what it is for, and its known limits and biases. Learn more

N

Noise schedule
The rule that sets how much noise is added at each step of the forward process. Learn more

O

Opt-out
A way for creators to ask that their work be removed from training data. Opt-in, by contrast, uses only work whose owners said yes. Learn more

P

Provenance
A record of where data or an image came from. For generated images, signed metadata (content credentials) can record that AI made them. Learn more

Q

Quantisation
Storing a model's weights as lower-precision numbers so it runs faster and uses less memory, with a small loss of quality. Learn more

R

Rebalancing / reweighting
Making under-represented examples count more during training, or adding more of them, so the model learns a fairer spread. Learn more
Representation bias
Some groups, places or styles appearing much more or less often in the data than in the world the model serves. Learn more
Reverse process
The learned process of removing noise step by step, used to generate images. Learn more

S

Sampling steps
The number of denoising passes used to generate one image. More steps is slower but can add detail. Learn more
Seed
The number that sets the random starting noise. The same seed, prompt and settings give the same image. Learn more
Skip connection
A link that copies feature maps from an encoder level straight to the matching decoder level, keeping fine detail. Learn more
Spectral normalisation
A limit on how sharply a discriminator's output can change, which keeps GAN training more stable. Learn more
Super-resolution (upscaling)
Enlarging an image while adding plausible detail, often with a GAN-based model. Learn more

T

Text encoder
The model that turns a prompt into an embedding the image network can use. Learn more
Throughput
How many images a system produces in a given time. Batching raises it, sometimes at the cost of latency. Learn more
Training data
The examples a model learns from: for image generators, very large sets of images with captions. Learn more
Transformer
A network built on attention layers. Some newer image generators use one in place of the U-Net. Learn more

U

U-Net
An encoder-decoder CNN with skip connections, whose output is the same size as its input. The usual denoiser in diffusion models. Learn more

V

Vanishing gradient
When the gradient a network learns from becomes almost zero, so its weights stop changing. In GANs, caused by an overconfident discriminator. Learn more
Variational autoencoder (VAE)
An encoder that maps images to a smooth latent distribution and a decoder that rebuilds them. Fast, but images can be blurry. Learn more

W

Wasserstein loss
A GAN loss that scores how far apart the real and generated distributions are, giving smoother gradients than "real or fake". Learn more
Watermark
A mark, visible or hidden in the pixels, that identifies an image as AI-generated or shows who made it. Learn more

Glossary

65 key terms, in plain English

Practice questions

Paper 1 Section B style, with full answers

Written by Spine in our own words. Spine is not affiliated with or endorsed by the International Baccalaureate. Every question, set of marking points and model answer here is our own, set in our own made-up company. Facts about the case study itself come from public guides and are marked “reported”. Last checked 2026-10-23.