2027 IB Computer Science Case Study Practice

What the 2027 case study requires

The May and November 2027 IB Computer Science case study is Generative AI for image creation: a diffuse vision. It places students in the context of Visionary Studios, a creative design company evaluating generative AI for advertising, concept art and digital media. The five original case study practice questions below cover definitions, explanations and evaluation, with model answers for both SL and HL.

Under the syllabus first assessed in 2027, the case study appears in Paper 1, Section B for both levels. SL students prepare the first two challenges, concerning diffusion models and ethical use of training datasets. HL students prepare those challenges plus GAN training and hybrid models.

These questions are practice material, not official IB questions or predictions. The model answers illustrate the depth, terminology and case-study application that strong responses should contain.

Paper 1 Section B at a glance

The official IB curriculum update confirms that Paper 1 combines Theme A with the case study. All Paper 1 questions are compulsory.

The official case study identifies these four research challenges:

1. Managing iterative denoising and the computational demands of diffusion models.
2. Addressing intellectual property and bias when using large training datasets.
3. Balancing generator and discriminator performance in GANs. This is HL only.
4. Evaluating combinations of VAEs, GANs, flow-based models and diffusion models. This is HL only.

Students should obtain the official booklet through their teacher or the IB Programme Resource Centre. A publicly accessible copy of the 2027 case-study booklet can also help you check its scenario, terminology and level-specific sections.

Five case study practice questions with model answers

Question 1: Defining core diffusion terminology

SL and HL:4 marks

Define the following terms in the context of generative image creation:

1. Noise injection [2]
2. Denoising [2]

Model answer

Noise injection is the introduction of random noise into image data. During diffusion-model training, noise is progressively added to training images so that the model can learn how different levels of corruption affect their features.

Denoising is the process of predicting and removing noise from an image. During generation, a trained neural denoiser repeatedly transforms an initial field of random noise into a coherent image that reflects patterns learned from the training dataset.

Why this earns marks: Each definition identifies the process and explains its role. Saying only that noise injection “adds noise” or denoising “removes noise” is unlikely to show sufficient contextual understanding.

Question 2: Explaining iterative denoising

SL and HL:4 marks

Explain why the iterative denoising process can help Visionary Studios create high-quality images but also create computational difficulties.

Model answer

A diffusion model starts with random noise and applies a neural denoiser over multiple stages. At each stage, the model predicts some of the noise and removes it, gradually reconstructing shapes, textures and other visual features learned during training. Repeated refinement can produce detailed, photorealistic and stylistically coherent images, which would be valuable for Visionary Studios’ commercial projects.

However, each denoising step requires another neural-network calculation. Producing one image may therefore require many sequential operations, increasing generation time, GPU use, energy consumption and cost. Visionary Studios could reduce this problem through more efficient sampling methods, compressed latent representations or suitable cloud hardware, but faster generation may involve a trade-off with quality or control.

The foundational DDPM research by Ho, Jain and Abbeel is useful supporting research because it explains image synthesis through diffusion and denoising. In an examination, name the underlying mechanism rather than merely claiming that diffusion is “slow.”

Question 3: Ethical and legal dataset management

SL and HL:6 marks

Discuss whether Visionary Studios should train an image-generation model using images collected from the open internet.

Model answer

Internet collection could provide a large and varied dataset, helping the model learn many objects, styles, cultures and visual contexts. Greater diversity may improve flexibility across advertising, concept art and digital-media assignments. Collection may also be quicker and cheaper than commissioning an entirely original dataset.

Nevertheless, public accessibility does not mean that an image is free of copyright or licensed for machine-learning use. A scraped dataset may contain protected artwork, personal data, biased representations or inaccurate labels. These problems could expose Visionary Studios to legal disputes, reputational damage and outputs that reproduce stereotypes or resemble protected material.

Visionary Studios should therefore not use indiscriminate scraping. It should build a curated dataset from original, licensed, public-domain or otherwise lawfully usable material; document provenance and permissions; test outputs for memorisation and bias; and provide a process for complaints and removal requests. A limited, carefully governed internet collection may be defensible where permissions and applicable law are checked, but unrestricted scraping creates excessive risk for a commercial studio.

This judgement is consistent with the NIST Generative AI Profile, which identifies intellectual-property and harmful-bias risks. Copyright rules differ by jurisdiction, so students should avoid claiming that every training use is automatically legal or illegal. The US Copyright Office AI reports demonstrate why the legal analysis depends on the circumstances and human contribution.

Question 4: GAN training stability

HL only:6 marks

Explain how the relationship between the generator and discriminator can create training instability in a generative adversarial network. Include mode collapse in your answer.

Model answer

A GAN contains a generator, which creates candidate images from random input, and a discriminator, which attempts to distinguish generated images from real training images. The generator improves by learning to fool the discriminator, while the discriminator improves at detecting artificial outputs. This creates an adversarial dynamic in which each network’s learning target changes as the other network changes.

If the discriminator becomes too accurate too quickly, the generator may receive weak or unhelpful learning signals and fail to improve. If the generator advances too quickly, it may repeatedly exploit a narrow weakness in the discriminator instead of learning the full training distribution. Careful adjustment of learning rates, architectures, update frequency and training data is therefore required to maintain balance.

Mode collapse occurs when the generator produces only a small range of similar outputs because those outputs successfully fool the discriminator. For Visionary Studios, this would reduce creative variety and make a GAN unsuitable for campaigns requiring many distinct concepts. Monitoring output diversity, modifying training objectives and using regularisation can reduce the risk, although these measures do not guarantee stable training.

Question 5: Evaluating the model choice

HL:12 marks. SL adaptation: evaluate diffusion models only for 6 marks.

Evaluate whether Visionary Studios should use a diffusion model, a GAN or a hybrid model for high-quality advertising images.

Model answer

A diffusion model is the strongest default where image quality, training stability and flexible conditioning are priorities. Its iterative refinement can generate detailed images from text, sketches or other conditions. The principal limitation is computational efficiency: repeated denoising can increase latency, hardware demand and operating cost.

A GAN can generate sharp images efficiently after training, making it attractive where rapid production matters. However, adversarial training is difficult to balance, and mode collapse can reduce variety. These weaknesses are significant for a studio that needs reliable output across different clients and campaigns.

A hybrid model could combine complementary strengths. For example, a VAE could provide an organised latent space for controllable editing, while diffusion produces the final high-quality image. A GAN component might improve sharpness or generation speed. The disadvantage is greater architectural complexity, more difficult debugging, higher development cost and potentially heavier computation.

Visionary Studios should therefore begin with a well-established conditional diffusion model supported by licensed, representative data and human review. A hybrid model is justified only when testing shows that diffusion alone cannot provide the required character consistency, editing control or speed. A standalone GAN would be less suitable as the general solution because its efficiency does not outweigh its instability and diversity risks, although it may remain useful for a narrow, well-tested task.

A strong evaluation does not merely list advantages and disadvantages. It weighs the alternatives against Visionary Studios’ stated criteria, including output quality, computational efficiency, training stability, flexibility and scalability, before reaching a conditional conclusion.

How to improve these model answers

Use a three-layer structure in every response:

• Technical knowledge: Explain the relevant process accurately.
• Case-study application: Connect it directly to Visionary Studios, its infrastructure or its clients.
• Judgement: For evaluative questions, weigh competing factors and reach a supported conclusion.

Command terms determine the required depth. Define requires a precise meaning, while explain requires connected reasoning showing how or why something happens. Discuss and evaluate require multiple perspectives, developed consequences and a justified conclusion.

Avoid memorising the model answers word for word. Instead, create a terminology table and practise reconstructing answers under timed conditions. RevisionDojo’s new-syllabus Computer Science Questionbank can provide broader exam preparation practice, while the Computer Science predicted papers can support full-paper timing.

Common mistakes in case study answers

• Writing generic AI answers: Refer to diffusion, denoising, dataset curation, GANs or hybrid architecture rather than discussing “AI” vaguely.
• Ignoring the level boundary: GAN and hybrid-model sections are HL only. SL students should concentrate on diffusion and ethical dataset use.
• Listing without developing: A point such as “diffusion is expensive” needs a cause and consequence.
• Treating law as universal: Copyright rules and exceptions vary across jurisdictions and circumstances.
• Giving an absolute conclusion: A model is suitable for a particular purpose, not universally superior.
• Relying only on the booklet: The case study is a research stimulus. Students should add relevant external research and examples.

For connected Theme A revision, use the A1 Computer Fundamentals questionbank and RevisionDojo’s guide to the most important Computer Science Paper 1 topics. The overview of changes in the new IB Computer Science syllabus is helpful if you are still distinguishing the 2027 course from the previous specification.

A practical revision method

1. Annotate each case-study section with its terminology, process, advantages, limitations and ethical implications.
2. Prepare SL challenges 1 and 2 first. HL students should then add GANs and hybrid architectures.
3. Answer one short question without notes, then compare it with the model answer.
4. Write evaluative answers using the sequence position, evidence for, evidence against, weighing, conclusion.
5. Ask Jojo AI to critique whether your reasoning is technically accurate and sufficiently connected to Visionary Studios.
6. Complete a timed Paper 1 practice session and record recurring weaknesses.

The central aim is not to predict exact wording. It is to become capable of applying researched knowledge to any reasonable question based on the four published challenges.

Conclusion

The 2027 case study assesses more than memorised definitions. Strong answers combine accurate computer science, direct application to Visionary Studios and balanced judgement about technical, ethical and commercial trade-offs. SL students need secure knowledge of diffusion models and dataset ethics, while HL students must also understand GAN training and hybrid models.

RevisionDojo can support this preparation through the Questionbank, predicted papers and Jojo AI feedback. Use these tools to practise retrieval and timed writing after you have studied the official booklet carefully.

Sources and referenced URLs

• Official IB Computer Science curriculum update
• 2027 Computer Science case-study booklet
• Denoising Diffusion Probabilistic Models research paper
• NIST Generative AI Risk Management Profile
• US Copyright Office artificial intelligence reports
• RevisionDojo new-syllabus Computer Science Questionbank
• RevisionDojo Computer Science predicted papers
• RevisionDojo A1 Computer Fundamentals Questionbank
• RevisionDojo guide to important Paper 1 topics
• RevisionDojo overview of the new Computer Science syllabus