How the case study is examined
| SL | HL | |
|---|---|---|
| Paper 1 | 1 hour 15 minutes, 50 marks, 35% of the grade | 2 hours, 80 marks, 40% of the grade |
| Section B (case study) | 12 marks: short-response questions | 24 marks: short- and extended-response questions |
| Based on | The booklet and its 2 challenges | The booklet and its 4 challenges |
| Structure | Shorter AO1/AO2 parts first, worth half the Section B marks, then one longer AO3 part worth the other half. The exact marks for each part may differ from the specimen paper (SL 2, 4, 6; HL 2, 2, 4, 4, 12), but this structure stays the same (IB FAQ). | |
| Timing | Sections A and B are done together, with no break. All questions are compulsory. At about 1.5 minutes per mark, spend about 20 minutes on Section B at SL and 40 minutes at HL. At SL the first two parts are short "Outline" answers: keep to the space given and save time for the last part. | |
The questions can test whether you can (from the IB Computer Science guide):
- show how the systems in the case study work, and the computational thinking behind them (AO1);
- apply concepts from the course in the context of the case study, and relate it to other scenarios (AO2);
- discuss the impacts and ethical issues, and evaluate, formulate and justify strategies using the booklet, your own research and new information given in the exam paper (AO3).
The booklet's cover says it is "required for" the paper 1 examinations, so expect to have a copy, but there is no time to learn it in the exam: know it well beforehand, and remember that your own research is only in your head. Expect the paper to add new details about Visionary Studios.
Where to focus your research: the booklet's last section, "Challenges faced", lists the issues to investigate, and they are different for SL and HL (IB FAQ). The IB publishes no reading list, so build your own sources around those challenges.
Answer technique
Marks tell you how much
- Roughly one clear, distinct point per mark
- "Outline" (2 marks): two short points; "Explain" (4 marks): points with reasons (because…, so…)
- See the command terms
Always use the scenario
- Write about Visionary Studios, its adverts, concept art and clients, not generative AI in general
- Use details from the question's new information
Use the booklet's terms
- Denoising, noise injection, DDPM, CNN, dataset curation, bias mitigation…
- HL: adversarial dynamic, mode collapse, latent space…
- Revise them on the terminology page
"Discuss" / "Evaluate" / "To what extent"
- Both sides (advantages and disadvantages, or several options)
- Linked to the studio's needs
- End with a justified conclusion or recommendation
The challenges: what you should be able to do
1. Iterative denoising and computational demands (SL and HL)
- Describe the forward (noising) and reverse (denoising) processes, and the role of the CNN.
- Explain why diffusion needs so much computation (steps × network size × resolution × number of images; training on huge datasets).
- Explain at least three ways to reduce it (fewer steps / better samplers / distillation; latent diffusion; lower resolution then upscaling; GPUs/TPUs, mixed precision, cloud; fine-tuning a pre-trained model) and the trade-off of each.
- Recommend a strategy for a given situation and justify it.
Notes: diffusion · Notes: computational demands · Lab: cost calculator
2. Ethics of large training datasets: intellectual property and bias (SL and HL)
- Explain how copyrighted material gets into training data and why that is a legal and ethical risk for the studio (with a real example such as Getty Images v Stability AI).
- Describe dataset curation steps (licensed or own data, opt-outs, records of sources, removing watermarks and personal images).
- Explain how bias in data leads to biased outputs, and describe bias-mitigation methods.
- Discuss transparency and AI disclosure, and evaluate the cost of doing things properly against the risks of not doing so.
3. Balancing the generator and discriminator in GANs HL
- Describe the roles of the generator, discriminator and noise vector, and the order of a training round.
- Explain what goes wrong when one network is much stronger (vanishing feedback, oscillation, mode collapse).
- Explain mitigation techniques (balancing learning rates and update ratios, label smoothing, Wasserstein loss, spectral normalisation, minibatch discrimination, progressive growing) and what each fixes.
4. How a hybrid model could balance VAEs, GANs, flow-based and diffusion models HL
- Compare and contrast the four families on quality, speed, stability, diversity, control and transparency.
- Describe a hybrid (e.g. latent diffusion = VAE + diffusion; diffusion + GAN distillation) and which weakness each component covers.
- Evaluate a hybrid for a given brief, including its costs (complexity, more to train and debug).
Practice questions
Written for this site in the style of Section B: they are not official IB questions, and the mark points are a guide. Answer on paper first, then open the mark points.
[2]SL + HL1. Outline the difference between conditional and unconditional image generation.
Mark points
- Conditional generation uses an input/condition (e.g. a text prompt, class label, sketch, segmentation map) to guide the output [1]
- Unconditional generation has no input other than random noise, so the output is based only on patterns learned from the training data [1]
[4]SL + HL2. Explain how a diffusion model generates a new image from noise.
Mark points
- Generation starts from an image of pure random noise (noise injection) [1]
- A neural network (CNN) predicts the noise present in the image… [1]
- …and part of it is removed, giving a slightly less noisy image [1]
- This is repeated over many steps until a clear image remains [1]
- The content depends on the training dataset / the condition such as a prompt [1]
Max 4.
[3]SL + HL3. Visionary Studios finds that generating images for a campaign takes too long. Describe three ways the computational cost of its diffusion model could be reduced.
Mark points
- Reduce the number of denoising steps, e.g. a faster sampler or a distilled model [1]
- Use latent diffusion: denoise a compressed representation instead of full-resolution pixels [1]
- Generate previews at low resolution and only upscale the chosen images [1]
- Run on more/faster GPUs or TPUs (e.g. rent cloud GPUs during busy periods) so the work finishes sooner [1]
- Use lower-precision (16-bit) arithmetic or batch several images per pass [1]
Any three; each must say what is done, not just "use better hardware" without detail. Extra GPUs reduce the time, not the amount of computation. Fine-tuning a pre-trained model reduces training cost, not generation time, so it does not answer this question.
[2]SL + HL4. Identify two ways in which the dataset used to train a diffusion model affects the images it produces.
Mark points
- It decides what content the model can produce (e.g. trained on animals → animals, cities → urban scenes) [1]
- Biases/imbalances in the data appear in (or are amplified in) the outputs [1]
- Copyrighted images or watermarks in the data can be reproduced/memorised [1]
- Image quality/style of the data sets the quality/style of outputs [1]
[6]SL + HL5. Visionary Studios plans to train its own model on images collected from the internet. Discuss the ethical and legal issues of this plan.
Mark points
- Many web images are copyrighted; using them without permission risks legal action (e.g. Getty Images v Stability AI) [1]
- Outputs may closely copy training images, watermarks or famous characters, exposing clients to claims [1]
- Artists' work is used without consent or payment and their style imitated: reputational damage with the creative community [1]
- Web data reflects social biases, leading to exclusionary or inaccurate representations in adverts [1]
- Images may contain personal data/faces without consent [1]
- Mitigations: licensed/own/public-domain data, opt-outs, dataset audits, record keeping, human review, disclosure [1–2]
- Counterpoint: curated data is expensive and smaller, which may reduce quality/variety; the law is still developing [1]
- Justified conclusion [1]
For full marks: both sides, linked to the studio, and a conclusion. Max 6.
[3]SL + HL6. Explain how an embedding-based approach could help Visionary Studios keep its mascot consistent across a series of adverts.
Mark points
- The mascot is represented as an embedding: a vector of numbers learned from reference images [1]
- The same embedding is given to the model with every prompt [1]
- So the mascot's features stay the same while the rest of the prompt changes the scene / without it each image starts from new noise and the character drifts [1]
[4]HL7. Explain the roles of the generator and the discriminator in a GAN.
Mark points
- The generator takes a D-dimensional noise vector and produces a synthetic image [1]
- The discriminator receives real images (training data) and fake images and predicts whether each is real or fake [1]
- The discriminator is trained to classify correctly / to get better at spotting fakes [1]
- The generator is trained using the discriminator's output to make fakes that are classified as real [1]
- They are trained together in competition (adversarial dynamic) [1]
Max 4.
[2]HL8. Define mode collapse.
Mark points
- A GAN training failure in which the generator produces a limited variety of outputs [1]
- …instead of the full diversity of the training dataset, e.g. the same few images repeatedly [1]
[4]HL9. The studio's discriminator quickly learns to reject every generated image, and the generator stops improving. Suggest two changes to the training process and explain how each helps.
Mark points
- Rebalance the learning rates / update ratio, e.g. lower the discriminator's learning rate or update it less often [1] so it does not overpower the generator and its feedback stays useful [1]
- Label smoothing (target 0.9 for real) [1] so the discriminator is less over-confident [1]
- Use a Wasserstein loss with gradient penalty / spectral normalisation [1] which gives smoother, still-useful feedback even when the discriminator is winning [1]
- Add noise to the discriminator's inputs / use a simpler discriminator [1] to slow its learning [1]
2 changes × (change + explanation).
[6]HL10. Evaluate the use of a hybrid model that combines a VAE with a diffusion model for producing high-resolution advertising images at Visionary Studios.
Mark points
- The VAE encoder compresses images into a latent space; diffusion runs in that smaller space; the decoder produces the final image [1]
- Advantage: far less computation per step than pixel-space diffusion, so faster/cheaper on the studio's infrastructure [1]
- Advantage: keeps diffusion's high quality, diversity and stable training [1]
- Advantage: the latent space supports control and editing / conditioning on prompts [1]
- Disadvantage: the VAE's compression can lose fine detail, e.g. small text or faces may be distorted [1]
- Disadvantage: still needs many denoising steps (slower than a GAN); more complex system with two models to train and maintain [1]
- Justified conclusion, e.g. suitable because quality and cost both matter for advertising [1]
Max 6; a top answer weighs both sides for this studio and concludes.