IB Computer Science HL · May/November 2027

Generative AI for image creation

The pre-seen case study follows a creative design company evaluating generative AI approaches for image production. HL Paper 1 Section B is worth 24 marks, so preparation combines technical understanding, independent research, application and balanced evaluation.

Course homeIA hub
Scenario

What the company is deciding

2 / 8

Image-generation approaches

Text-to-image, conditional/image-to-image and unconditional generation can serve different creative goals.

Candidate architectures

Diffusion models are central. HL also studies GANs and hybrid models in depth, while VAEs and flow-based models need broad understanding for hybrid-model reasoning.

Decision criteria

Output quality, computational efficiency, training stability, flexibility/scalability, character consistency and style adherence.

Ethical/legal factors

Dataset curation, intellectual property, bias/fairness and transparent disclosure of AI-generated content.

HL challenges

Four research questions drive the preparation

3 / 8

1 · Diffusion compute

Explain iterative denoising and evaluate the computational demands of producing high-quality images.

2 · Training data

Evaluate intellectual-property and bias risks in large training datasets and strategies for mitigation.

3 · GAN balance

Explain why generator and discriminator performance must remain balanced enough for useful adversarial training.

4 · Hybrid trade-offs

Evaluate how a hybrid system could combine diffusion, GAN, VAE and flow-based strengths while reducing weaknesses.

Research ruleBuild evidence around these challenges. The examination can also provide new stimulus, so learn to transfer your research rather than memorize a single essay.
Diffusion

From random noise to a coherent image

4 / 8
Training idea
Learn how progressively corrupted/noisy images can be related to cleaner representations.
Generation
Start from noise and repeatedly apply a learned denoising model until structured visual features emerge.
Trade-off
Repeated inference steps can produce high-quality results but increase computation, memory traffic, latency, energy use and infrastructure requirements.
Hardware linkUse CPU/GPU architecture, parallelism, memory access and deployment scale to make the “computational demand” discussion concrete.
GANs + hybrid models

Competing networks and complementary architectures

5 / 8

GAN training

The generator creates candidates; the discriminator judges real versus generated examples. If one becomes too strong relative to the other, useful learning can stall or become unstable.

Mode collapse is one failure mode: the generator produces too little variety.

Hybrid reasoning

Do not treat “hybrid” as automatically better. Evaluate what each component contributes, what complexity/cost it adds, and whether the combined system actually improves quality, control, stability, transparency or reproducibility.

Depth boundaryHL should know diffusion, GANs and hybrid models deeply enough for the challenge questions. VAE and flow-based models need broad understanding sufficient to evaluate their role inside a hybrid design.
Terminology

Case-study vocabulary bank

6 / 8

Diffusion model

Generates images by learning to reverse a gradual noise process through repeated denoising steps.

Denoising

Removing noise step by step so a noisy representation becomes a coherent image.

DDPM

A probabilistic diffusion framework that formalizes the gradual addition and learned removal of noise.

CNN

A neural-network architecture that learns spatial image features; commonly used in image processing and denoising components.

Conditional generation

Generation controlled by input such as text, a class label, a sketch, a segmentation map or another image.

Unconditional generation

Generation without an external condition, relying on patterns learned from the training data.

GAN

A generative system trained through competition between a generator and a discriminator.

Generator

Creates synthetic candidate images from random input.

Discriminator

Tries to distinguish real training examples from generated examples.

Adversarial dynamic

The competitive training relationship that drives generator and discriminator improvement.

Mode collapse

A GAN failure mode where the generator produces too little variety.

VAE

A generative model that learns a structured latent representation and can sample from it to produce new data.

Latent space

A compressed numerical representation in which learned features are encoded.

Flow-based model

A model based on an invertible mapping between data and a simpler probability distribution.

Hybrid model

Combines different generative approaches so one model can compensate for limitations of another.

Dataset curation

Selecting and checking training data for relevance, quality, legality and representational balance.

Bias mitigation

Steps taken to reduce systematic unfairness in data, training or generated outputs.

Training stability

How reliably training converges toward useful behaviour instead of collapse, divergence or severe oscillation.

Character consistency

Keeping a recurring subject recognizably consistent across multiple generated images.

Embedding-based approach

Representing concepts/features as numerical vectors so identity, similarity or style can be controlled.

Text-to-image generation

Creating images from natural-language descriptions.

Image-to-image translation

Transforming one image representation into another, such as sketch-to-render.

Noise injection

Adding random noise to data as part of the diffusion training/generation framework.

Segmentation map

A labelled spatial map of image regions that can be used as a generation condition.

Image generation

Creating new visual outputs using a generative model.

D-dimensional noise vector

A D-value random vector used as input to a generative model such as a GAN.

Course research timeline

The case study is a strand, not a final-week topic

7 / 8
L10
Formal launch: scenario, four HL challenges, first diffusion research note.
L20
GAN generator/discriminator balance and training stability.
L29
CPU/memory connection to iterative image-generation workloads.
L37
CPU vs GPU, parallelism, memory access and deployment trade-offs.
L42
Networking/cloud/security considerations for deploying generative AI services.
L49
Dataset scale, curation, intellectual property and bias research.
L50–L54
Deep integration with machine-learning syllabus content and terminology.
L55
Final synthesis and Paper 1 Section B practice — not first exposure.
Paper 1 Section B

Turn research into an exam response

8 / 8

Short parts

Retrieve precise definitions, mechanisms and direct applications quickly. Use the vocabulary of the case study accurately.

AO3 extended response

Build a balanced technical argument: identify criteria, analyse trade-offs, use research evidence, respond to any new stimulus, and reach a justified conclusion tied to the scenario.

Time planningHL students should be prepared to spend roughly 40 minutes on the 24-mark case-study section, while adapting to the exact mark breakdown on the live paper.

Evidence bank checklist