2027 Case Study: Generative AI for image creation
For Paper 1, Section B. May and November 2027. Sat by both SL and HL. Primary syllabus home: Theme A4 (machine learning), with A1 fundamentals (the CNN denoiser, computational resources) and the cross-theme ethics strand.
The 2027 case study, “Generative AI for image creation, a diffuse vision,” asks you to step into a design company that is deciding how to use generative AI. To answer well you need to understand how the main image-generation techniques work, where each is strong and weak, how to compare them, and what the ethical and legal risks are. These pages take you through all of that, marking clearly what is for everyone and what is HL only.
Table of Contents
- The scenario
- Why generative AI, and what is being judged
- How it is examined
- What you need to know, by level
- How to use these pages
The scenario
The booklet centres on Visionary Studios, a creative design company. It produces images for advertising campaigns, concept art, and digital media, and it is evaluating generative AI to make that production faster and more flexible without losing quality. It is weighing up three families of model: diffusion models, GANs, and hybrid models, each with different strengths, costs, and risks.
Everything you write in the exam should come back to this company and its goals. A general point about “AI” earns little; the same point tied to Visionary Studios producing a consistent brand mascot across a campaign, on a real compute budget, is what the mark scheme rewards.
Why generative AI, and what is being judged
Visionary Studios is not asking “is AI good.” It is asking a practical engineering and business question: which model should we use, and what do we have to manage to use it responsibly? That means weighing image quality against the cost of running the model, the reliability of training, the ability to keep a character or style consistent across many images, and the legal and ethical exposure that comes with training on large image datasets.
How it is examined
Section B is one question, split into parts. The shape follows the official specimen:
| SL (12 marks) | HL (24 marks) | |
|---|---|---|
| Short, point-marked parts | About 2 (Identify, Outline) | About 4 (Identify, Outline) |
| Extended response | One, 6 marks, marked against a 4-band scale that tops out at Competent | One, 12 marks, marked against a 5-band scale reaching Proficient |
| Challenges in scope | The 2 SL+HL challenges | All 4 challenges |
The short parts test whether you can state a precise point and expand it. The extended response tests whether you can argue more than one side, anchor it to the scenario, use the terminology accurately, bring in your own research, and reach a justified conclusion. The exam prep page works through both with marked examples.
What you need to know, by level
This is the single most important thing to get right, because the booklet sets different expectations for SL and HL. Learn your level’s depth and do not waste time beyond it.
| Technique | SL | HL |
|---|---|---|
| Diffusion models (denoising, DDPM, CNN denoiser) | In depth | In depth |
| Image generation techniques (text-to-image, conditional, unconditional) | In depth | In depth |
| Evaluating models and the ethics of training data | In depth | In depth |
| GANs (generator, discriminator, mode collapse) | Not required | In depth |
| Hybrid models (why combine approaches) | Not required | In depth |
| VAEs and flow-based models | Not required | Broad understanding only |
SL students: GANs and hybrid models are not in your scope. The booklet states you are not expected to know them, and SL exam questions sit only on the two SL+HL challenges. The GANs and hybrid pages below are marked HL only. Use the HL / SL toggle at the top of any page to hide HL-only sections.
How to use these pages
The booklet is a springboard, not a reading. Treat the 2027 case study as a small research project: define every term in your own words, find real tools and real debates, and practise arguing both sides of each challenge. The pages below follow the booklet’s own order.
For everyone (SL and HL):
- Image generation techniques - text-to-image, conditional, and unconditional generation, and what each is for.
- Diffusion models - the core technique: noise, denoising, DDPM, and the CNN denoiser.
- Evaluating generative AI models - the five factors Visionary Studios uses to compare options.
- Ethics and law - dataset curation and IP, bias and fairness, transparency and disclosure.
- Glossary trainer - every term you must be able to define and apply.
- Challenges and exam prep - the challenges, the markbands, the answer scaffold, and worked questions.
HL only (in addition to the above):
- GANs - the generator and discriminator, the adversarial dynamic, and mode collapse.
- Hybrid models - VAEs, flow-based models, latent space, and why you would combine approaches.
EduCS.me is an independent teaching resource, currently aligned with the IB CS (2027) syllabus and not affiliated with the IB. Always work from your own clean copy of the official booklet.