The 2027 IB Computer Science case study four challenges identify the main problems Visionary Studios must investigate when considering generative AI for image creation. The first two challenges, diffusion-model efficiency and responsible use of training data, apply to SL and HL. The final two, balancing GAN networks and evaluating hybrid models, are HL only.
These challenges are testable because they require more than memorized definitions. Each presents a decision involving competing technical, financial, creative, or ethical factors. That makes them suitable for questions asking students to explain a problem, apply knowledge to Visionary Studios, compare possible approaches, and justify a conclusion.
What the four challenges cover
The pre-seen case study, Generative AI for image creation: a diffuse vision, places students in the context of Visionary Studios, a creative design company producing advertising images, concept art, and digital media. The company is assessing generative AI models that could streamline production while maintaining quality and creative control.
The case study assigns the challenges by level as follows:
| Challenge | Central issue | SL | HL |
|---|---|---|---|
| Diffusion-model denoising | Photorealistic quality versus processing time and computational demand | Yes | Yes |
| Training-data ethics | Useful large datasets versus intellectual property and bias risks | Yes | Yes |
| GAN balancing | A capable generator versus a discriminator that provides useful feedback | No | Yes |
| Hybrid-model evaluation | Combining model strengths versus added complexity and cost | No | Yes |
The booklet explicitly states that SL students are not expected to know about GANs or hybrid models. HL students should study diffusion models, GANs, and hybrid models in depth, while only a broad understanding of variational autoencoders and flow-based models is expected.
Under the curriculum first assessed in 2027, the case study appears in Paper 1, Section B for both levels. This replaces the older arrangement in which a pre-seen case study was associated with a separate HL Paper 3. The IB's curriculum update confirms that Paper 3 has been removed and that the case study is now assessed in Paper 1.
Challenge 1: Managing diffusion-model denoising
A diffusion model begins image generation with random noise. It repeatedly predicts and removes part of that noise until a recognizable image emerges. Each denoising step moves the image closer to a result matching patterns learned from the training data and, in a conditioned system, the user's prompt.
The challenge is that iterative denoising requires repeated neural-network calculations. More steps may improve detail or consistency, but they also increase generation time, hardware use, energy consumption, and operational cost. Producing one attractive image slowly may be acceptable for concept art, while generating hundreds of advertising variations under a deadline creates a different requirement.
Possible responses include:
- reducing the number of sampling steps
- using optimized sampling schedules or faster solvers
- generating in a compressed latent space rather than directly in pixel space
- applying pruning, caching, quantization, or model distillation
- using more capable GPUs or distributed processing
- accepting slightly lower resolution for early drafts
Each response creates a trade-off. Fewer steps may improve speed but reduce fine detail; stronger hardware shortens waiting time but raises financial and environmental costs; latent processing improves efficiency but adds system complexity.
This challenge is testable because students can be asked to evaluate whether photorealism justifies the computational demand. A strong answer would distinguish training cost from inference cost, apply the issue to Visionary Studios, and reach a conditional judgment. For example, high-quality diffusion generation may be appropriate for final campaign images, while a faster, lower-quality configuration may be preferable during rapid prototyping.
Challenge 2: Intellectual property and bias in training data
Generative image models require large collections of examples from which they learn statistical patterns. The challenge is not simply obtaining enough images. Visionary Studios must consider whether those images were collected and used lawfully, whether creators gave permission, and whether the dataset represents people and cultures fairly.
Intellectual property concerns may include copyright ownership, licensing conditions, attribution, dataset provenance, and outputs that reproduce protected elements too closely. Copyright rules vary between jurisdictions, so students should avoid claiming that all training on copyrighted work is automatically legal or automatically illegal. The US Copyright Office, for example, describes generative-AI training as a fact-specific issue in which the source, purpose, works used, and safeguards can matter.
Bias occurs when the data, labels, collection process, model, or deployment context systematically favors some outcomes over others. If a dataset overrepresents particular genders, skin tones, professions, regions, or artistic traditions, generated advertising images may reproduce stereotypes or exclude groups. NIST also emphasizes that AI bias can be systemic, computational and statistical, or human-cognitive, so representativeness alone does not solve every fairness problem.
Practical mitigations include:
- documenting where training images came from
- licensing content or using approved and public-domain collections
- allowing creators to opt in or opt out where appropriate
- auditing datasets for duplication and representation gaps
- testing outputs with prompts involving different demographic groups
- using human review before commercial publication
- recording known limitations and responding to complaints
The trade-off is between dataset scale and responsible governance. Restricting training data may increase cost or reduce coverage, but using poorly documented material can expose the company to legal disputes, reputational harm, and discriminatory outputs. This is highly testable because an exam question can ask students to balance stakeholders: Visionary Studios, artists, clients, depicted communities, model developers, and audiences.
Students can strengthen this area through the RevisionDojo A4.4 ethical considerations questionbank, especially by practising answers that connect a safeguard to a specific risk rather than listing generic ethical principles.
Challenge 3: Balancing GAN generator and discriminator performance
This challenge is HL only. A generative adversarial network contains two neural networks trained in competition:
- The generator converts a random input vector into a synthetic image.
- The discriminator examines real and generated images and predicts which are genuine.
The generator improves by learning to fool the discriminator. The discriminator improves by becoming more accurate at identifying generated images. Ideally, their competition supplies a useful learning signal and leads to realistic outputs.
The difficulty is that one network can become too capable relative to the other. If the discriminator becomes too accurate too early, the generator may receive weak gradients and stop improving effectively. If the generator exploits a weakness in the discriminator, it may repeatedly produce a narrow range of successful outputs.
That second failure is called mode collapse. For Visionary Studios, a GAN might create sharp images but repeatedly generate similar faces, layouts, or visual styles. Such output could look realistic individually while failing the company's need for variety.
Possible mitigation strategies include adjusting learning rates or update frequency, changing loss functions, using regularization, applying minibatch-based methods, monitoring output diversity, and improving the architecture or training data. No technique guarantees perfect balance, which is precisely why this is an evaluative challenge.
An exam question could ask whether GANs are suitable for varied campaign images. A strong response would weigh sharp output and potentially fast generation against unstable training, difficult evaluation, and mode collapse. HL students can reinforce the underlying machine-learning concepts with the A4.3 machine-learning approaches questionbank.
Challenge 4: Evaluating a hybrid generative model
The fourth challenge is also HL only. A hybrid model combines techniques so that one component compensates for another's weakness. The case study names VAEs, GANs, flow-based models, and diffusion models.
| Model family | Useful strength | Important limitation |
|---|---|---|
| VAE | Structured latent representation and useful attribute manipulation | Outputs can be less sharp |
| GAN | Sharp, realistic images and efficient generation after training | Unstable training and mode collapse |
| Flow-based model | Exact likelihood calculation and reversible transformations | Architectural constraints and substantial computational or memory demands |
| Diffusion model | High-quality, diverse generation and comparatively stable training | Slow, computationally intensive iterative sampling |
A hybrid might use a VAE encoder to compress an image into a latent representation, diffusion to generate or refine that representation, and an adversarial objective to encourage sharper output. However, combining components does not automatically remove their weaknesses. It can introduce more hyperparameters, competing objectives, debugging difficulty, longer development time, and greater hardware requirements.
This challenge is testable because it requires criteria-based evaluation, not a claim that hybrid is always better. Visionary Studios might prioritize visual fidelity, controllability, diversity, generation speed, training stability, cost, or ease of maintenance. The most suitable combination depends on which criteria matter most for the intended workflow.
A strong conclusion might recommend a hybrid for final high-value work if additional complexity produces measurable improvements in control and quality. For frequent low-cost drafts, a simpler model may be more maintainable. The key is to compare benefits against the engineering burden rather than describing each model separately.
Why these are testable case study topics
The four challenges effectively define the case study's research territory, but students should not assume that the exam will copy each bullet word for word. Questions may introduce new stimulus material or a particular production requirement and ask students to apply their research.
Each challenge supports several assessment skills:
- Knowledge: defining denoising, bias, mode collapse, or latent space
- Application: relating a concept to Visionary Studios
- Analysis: tracing how a technical choice creates a consequence
- Evaluation: weighing advantages and disadvantages against stated criteria
- Justification: selecting an approach and defending it with evidence
For an essay-style response, use a simple structure: establish the decision, identify the relevant criteria, explain benefits, explain limitations, compare realistic alternatives, and give a justified conclusion. Avoid writing two disconnected lists of advantages and disadvantages. Every paragraph should show why the point matters in this scenario.
How to prepare efficiently
Create one revision sheet for each challenge containing the mechanism, central trade-off, two or three mitigations, and a scenario-specific judgment. SL students need sheets for challenges 1 and 2; HL students need all four.
Then practise changing the recommendation when the scenario criteria change. A model suitable for a final photorealistic advertisement may not be suitable for rapid brainstorming, limited hardware, or a project requiring extensive style diversity.
Use RevisionDojo's 2027 Computer Science questionbank to practise command terms, and use the A4 machine-learning flashcards for precise terminology. After writing a response, Jojo AI can help identify unsupported judgments or places where an explanation has not been connected to Visionary Studios. Closer to the examination, Computer Science predicted papers can provide timed practice, but predictions should never replace preparation across the full published scope.
Conclusion
The four challenges are testable because each turns technical knowledge into a decision. Diffusion models trade image quality against computational demand; large datasets trade coverage against intellectual-property and bias risks; GANs trade sharp output against unstable adversarial training; and hybrids trade complementary strengths against greater complexity.
SL students should prepare the first two challenges, while HL students must prepare all four and study GANs and hybrid models in greater depth. RevisionDojo's Questionbank, Flashcards, predicted papers, and Jojo AI are most useful when used to practise applied explanations and justified conclusions rather than memorize one fixed response.
Sources and referenced URLs
- IB Computer Science curriculum update
- 2027 Computer Science case study booklet
- Denoising Diffusion Probabilistic Models
- NIST Generative Artificial Intelligence Profile
- US Copyright Office report on generative AI training
- Google Developers guide to common GAN problems
- RevisionDojo Computer Science 2027 questionbank
- RevisionDojo A4.3 machine-learning approaches questionbank
- RevisionDojo A4.4 ethical considerations questionbank
- RevisionDojo A4 machine-learning flashcards
- RevisionDojo Computer Science predicted papers
