START HERE | ACTIVE REVISION
Cover the explanations and try to explain each heading aloud or in your book. Then check your explanation, correct it and attempt the recall prompt. Use the Flashcards tab to select weak terms and build a revision list.
Sections 01–04 and 07–08 cover shared SL and HL content. Sections 05 and 06 are HL only . The level selector filters both the summaries and flashcards. “Shared SL & HL” means the common material; “HL only” isolates additional content.
These independent notes accompany the official 2027 booklet. All pilot figures below are fictional classroom examples, not booklet findings.
01 | THE STUDIO AND THE BASIC IDEAS
Visionary Studios is considering image generation for advertising, concept art and digital media. Judge a system against the company’s intended work: useful images, relevant style, consistency, resources and responsible production.
Generative AI learns patterns from training examples and uses them to produce content. During training , parameters are adjusted using examples and a learning objective. During ordinary inference , a trained model generates a result without retraining for each request.
Learning patterns does not guarantee originality, accuracy or fairness. A plausible image is not evidence that a depicted event happened. Also distinguish capability from suitability: producing one impressive image does not show dependable campaign production.
Recall challenge: Explain the difference between training and inference, then give one studio consequence.
Training changes learned parameters and can require substantial resources. Inference uses those parameters to generate candidates; its resource cost and output quality still need evaluation.
02 | WAYS TO GUIDE GENERATION
Text-to-image uses a text prompt to guide the result. A useful prompt can specify subject, setting and visual requirements, but wording alone does not guarantee precise control.
Conditional generation uses additional information. A class label can guide category; a sketch can guide layout; a segmentation map identifies regions by category. These inputs constrain different aspects of the task and are not interchangeable.
Image-to-image translation transforms or generates an image using an input image, for example changing visual appearance while attempting to preserve useful structure. Unconditional generation samples without a specified conditioning input; it still depends on training and model design.
Synthetic data can help create examples, but must be checked for usefulness and bias. Generated does not mean representative or automatically suitable for training another system.
Recall challenge: Which input would guide object-region layout more explicitly than a short class label?
A segmentation map can assign categories to regions, giving spatial guidance. A class label usually indicates category without specifying detailed layout.
03 | DIFFUSION AND ITERATIVE DENOISING
In a simplified diffusion training account, noise is added to training examples and a network learns a denoising-related task at different noise levels. Learning may involve predicting added noise or another related target, depending on the formulation.
Generation begins from a noisy sample. A sampling procedure repeatedly uses network predictions to update the sample towards image structure, with conditioning where applicable. This is not simply retrieving a hidden original image and is not retraining on each prompt.
A CNN can learn spatial features and serve within a denoising architecture. DDPM names a probabilistic diffusion framework; it is not another name for a CNN. A network architecture and a generation framework describe different roles.
More sampling steps can increase computation, but quality does not improve without limit. Evaluate the actual model, sampler and settings. In latent diffusion, iterative generation occurs in a learned representation and a decoder produces the displayable image.
Recall challenge: Why is “diffusion removes noise until it finds the original training image” misleading?
Sampling constructs an output through learned updates; it does not ordinarily identify a particular original to retrieve. Memorisation can still be a concern, so generation is not a guarantee against reproduction.
04 | QUALITY, CONSISTENCY AND COMPUTATION
Quality includes suitability, requested detail and absence of unacceptable artefacts. Style adherence concerns the requested appearance. Character consistency means preserving relevant identity across images while allowing intended changes such as pose.
Embeddings are numerical representations that can help condition generation or assess similarity. A similarity score is evidence under a method, not an automatic proof that a character meets every campaign requirement.
Measure total production effort : setup, retries, review and editing as well as individual generation. Training stability concerns reliable learning; computational efficiency concerns resources used for the useful result. Neither is established by visual quality alone.
Scalability concerns handling increased demand under stated conditions. Brand requirements can make precise details essential even when an output is attractive. A tool that works in a demonstration may not meet simultaneous production demand.
Recall challenge: Why can a fast generator still produce slow client delivery?
Rejected candidates and corrective editing add time before an asset is suitable and approved. End-to-end measurements reveal this effort.
05 | GANS AND ADVERSARIAL TRAINING | HL ONLY
A generator maps a noise vector, and conditioning where applicable, to a candidate. A discriminator learns to distinguish training examples from generated examples. Its feedback contributes to generator learning through the training objective.
Both networks are updated during adversarial training. The discriminator is not a human art judge, and its score does not establish client suitability. Ordinary trained-generator inference need not run the discriminator.
Mode collapse occurs when the generator produces a limited range of outputs instead of covering useful diversity. Training instability is broader: updates may fail to produce reliable improvement. Balance, data quality, objectives and regularisation can be investigated, but no single technique guarantees a cure.
GANs can have efficient generation, while diffusion commonly uses iterative sampling. Avoid declaring universal winners: compare actual trained systems under equivalent tasks, quality conditions and resource definitions.
Recall challenge: Why is a good discriminator score not sufficient for the studio?
The training signal addresses its learned discrimination task, not every brief requirement, identity feature, portrayal issue or client approval condition.
06 | LATENT MODELS AND HYBRIDS | HL ONLY
A VAE encoder learns a distribution for an input’s latent representation; a decoder reconstructs from sampled latent values. Reconstruction and regularisation encourage a usable latent space. Sampling a latent value for generation differs from reconstructing a particular input.
A normalizing flow uses invertible transformations between a simple distribution and a modelled data distribution, with the mathematical structure supporting density calculation. Reversibility does not reveal the real-world provenance of an image.
A hybrid combines complementary components or objectives. Latent diffusion can use an autoencoder with diffusion in latent space; other combinations may add adversarial or flow-related components. Explain the role of each part rather than assume every hybrid has the same architecture.
Evaluate whether added quality or control outweighs integration, setup and maintenance effort. Compare with appropriate simpler baselines and test which component contributes useful improvement. Broad conceptual understanding is the focus here, not deriving VAE or flow mathematics.
Recall challenge: Does combining models guarantee that all their weaknesses disappear?
No. Components can add complexity and new failure points. A representative pilot and component comparisons are needed to establish useful benefits.
07 | SOURCES, FAIRNESS AND TRUST
Dataset curation includes deliberate selection, checking and documentation. Keep provenance and relevant licence evidence. Public availability does not itself establish permission for every activity; attribution does not replace required permission. Applicable rules and contracts need context-specific review.
Separate input-use questions, potential interference with others’ rights and rights claimed in outputs. A provider’s commercial-use wording does not automatically establish exclusive copyright everywhere.
Bias can enter through coverage, captions, training, prompts and human selection. Examine who appears and how they are portrayed. Equal counts can hide stereotyped roles. Mitigate identified causes, test outputs and acknowledge residual concerns.
Transparency explains the AI contribution, completed checks and limits. Labels and metadata help communicate or record origin but do not prove factual truth. Clear disclosure supports informed decisions; it does not settle quality, fairness or rights on its own.
Recall challenge: Name a safeguard for sources, fairness and disclosure, and explain why they are distinct.
Keep source evidence; review representation and roles; explain generated contributions accurately. Each addresses a different issue, so success in one does not establish success in the others.
08 | TEST, RECOMMEND AND WRITE
Turn needs into criteria, measures and justified proposed conditions. Link each to a test and result. Distinguish booklet facts, justified inferences and classroom assumptions. Essential conditions are decision gates; preference scores should not silently override them.
For fair comparisons, define whether you test models or complete workflows. Use equivalent requirements, representative briefs and stated effort budgets. Record versions, settings and all attempts; use repeated outputs and a common rubric. Matching seed numbers across tools does not guarantee equivalent random inputs.
Where practicable, hide tool labels from reviewers and score independently. Inspect results by task as well as overall. Small tests support bounded conclusions, not universal superiority.
Recommend a specific action, cite relevant evidence, explain the company consequence, weigh an alternative and state important conditions. For writing, identify names items; define gives precise meaning; describe sets out features or stages; explain develops why or how; discuss develops relevant perspectives and a considered synthesis.
From 2027 the case study is assessed in Paper 1 at SL and HL. The 2, 2, 4, 4, 6 and 12-mark set in lesson 08.4 is classroom practice, not a guaranteed new paper structure. Use current official specimens and instructions.
Recall challenge: What makes a recommendation case-specific?
It connects accurate technical evidence to the studio’s needs and consequences, then justifies an action under relevant conditions rather than naming a favourite model.
WORKED RECALL | CALCULATIONS AND INTERPRETATION
| Measure | A | B |
|---|---|---|
| Accepted | 24 | 30 |
| Total time | 240 minutes | 330 minutes |
Calculate acceptance rates, their difference and time per accepted asset.
A: 24 ÷ 40 × 100 = 60%. B: 30 ÷ 40 × 100 = 75%. B is 15 percentage points higher. A takes 240 ÷ 24 = 10 minutes per accepted asset; B takes 330 ÷ 30 = 11 minutes. A is quicker on this measure, while B has a higher accepted proportion. Neither measure alone establishes the best overall option.
A baseline is 30 minutes per asset. Calculate a proposed 20% reduction target.
30 × 0.20 = 6 minutes saved. Target: 30 − 6 = 24 minutes or less. This is a proposed classroom condition, not a booklet requirement.
Setup takes 120 staff-minutes and savings are 2 staff-minutes per asset. Find the simplified time break-even point.
120 ÷ 2 = 60 assets. This assumes constant savings and ignores other differences. It is not a financial profit calculation.
COMMON MISTAKES | CORRECT THE CLAIM
| Mistake | Correction |
|---|---|
| Every prompt retrains the model. | Ordinary inference uses learned parameters. |
| More sampling steps always improve quality. | Evaluate the sampler, model and settings; gains are not unlimited. |
| A CNN and DDPM are the same thing. | Architecture and diffusion framework describe different roles. |
| Publicly accessible images are unrestricted. | Access alone does not establish a suitable basis for intended use. |
| Equal counts prove fairness. | Portrayal, context and roles also matter. |
| Higher acceptance proves higher quality among accepted images. | These are different measures. |
| A disclosure resolves every concern. | It communicates contribution; other checks remain. |
| Adding the company name makes an answer applied. | Explain relevant company consequences. |
FINAL CHECK | CAN YOU CONNECT THE IDEAS?
Without notes, explain one generation method, one consistency challenge, one fair comparison and one responsible-use decision. HL students should also compare adversarial generation with diffusion and explain the purpose of a hybrid.
Then write a short recommendation for the studio using a stated purpose, evidence, an alternative and a condition. Return to the relevant lesson if you cannot explain the mechanism or company consequence clearly.
Revisit worked exam answers · End of case study review · Official IB assessment update