The 2027 IB Computer Science case study, Generative AI for image creation: a diffuse vision, is assessed in Paper 1, Section B for both May and November candidates. The most damaging case study terminology mistakes usually involve treating related terms as synonyms, especially conditional and unconditional generation, noise injection and denoising, or generator and discriminator.
This guide corrects ten commonly confused terms through side-by-side definitions. It also identifies which distinctions matter at SL, which require additional HL knowledge, and how to use each term accurately in an exam response.
Why terminology matters in the 2027 case study
Under the curriculum first assessed in 2027, the case study is included in Paper 1 for both levels. The IB's Computer Science curriculum update confirms that the former HL Paper 3 has been removed and that the case study is now assessed within Paper 1 at SL and HL.
The case study examines Visionary Studios, a design company evaluating generative AI for advertising, concept art, and digital media. Students must connect technical knowledge to practical decisions involving image quality, computational demands, consistency, intellectual property, and bias.
Knowing a definition is therefore only the first step. A strong answer must distinguish the term from nearby concepts and apply it to the company's situation.
| Content area | SL | HL |
|---|---|---|
| Conditional and unconditional generation | Required | Required |
| Diffusion models, noise injection, and denoising | Required | Required |
| DDPM principles | Required | Required |
| Generator and discriminator in GANs | Not part of the SL-only challenges | Required |
| Mode collapse and adversarial training problems | Not part of the SL-only challenges | Required |
The distinction matters because HL students prepare all four case study challenges, while SL students focus on the first two. SL students should not spend disproportionate revision time on detailed GAN or hybrid-model architecture.
1 and 2: Conditional vs unconditional generation
| Conditional generation | Unconditional generation |
|---|---|
| Generates an image using an additional input that constrains what should be produced. | Generates an image without a user-supplied condition specifying its content. |
| The condition could be a class label, prompt, sketch, segmentation map, or source image. | The model samples from patterns learned across its training data. |
| Appropriate when Visionary Studios needs an output matching a brief. | Appropriate for open-ended exploration or synthetic data creation. |
The common confusion is that unconditional does not mean random in the sense of having learned nothing. The output is still governed by the model's learned probability distribution, but the user does not specify a category, layout, or other guiding condition.
Text-to-image generation deserves particular care. The case study presents it alongside conditional and unconditional techniques, but technically a text prompt acts as a condition because it guides the output. In an exam, describe text-to-image as a specific method in which textual information conditions image generation, rather than claiming that every conditional system must use text.
For example, generating any plausible abstract design without a brief is unconditional. Generating “a sustainable glass office beside a river at sunset” is conditional because the prompt constrains the subject, setting, and lighting.
3 and 4: Noise injection vs denoising
| Noise injection | Denoising |
|---|---|
| Introduces random noise into data or uses random noise as the starting state for generation. | Predicts and removes noise step by step to recover or construct coherent features. |
| During diffusion training, controlled noise is added to real images. | The neural denoiser learns how to reverse stages of corruption. |
| During generation, the model begins with sampled noise. | Repeated denoising transforms that noise into a new image. |
Students often describe diffusion as simply “removing noise from a photograph.” That wording implies the model is restoring one known original image. In generation, there is no hidden original image waiting underneath the noise; the model produces a new sample using patterns learned during training.
The sequence should be explained precisely:
- During the forward process used for training, increasing amounts of noise corrupt training images.
- A neural network learns to predict the noise or the information needed to reverse that corruption.
- During generation, the system starts with random noise.
- Iterative denoising gradually creates a coherent new image.
The original DDPM research paper formalizes diffusion as a process whose learned transitions reverse gradual corruption. This repeated processing helps produce high-quality images, but it also explains the substantial computational demand highlighted in the case study.
5 and 6: Diffusion model vs DDPM
| Diffusion model | Denoising diffusion probabilistic model |
|---|---|
| A broad family of generative models that learns to reverse a gradual noising process. | A particular probabilistic formulation of diffusion, abbreviated DDPM. |
| Describes the overall model category. | Formalizes forward noising and learned reverse transitions over multiple time steps. |
| May include different architectures, sampling methods, and optimizations. | Provides the foundational process emphasized in the case study. |
A DDPM is a diffusion model, but the terms are not perfectly interchangeable. “Diffusion model” names the wider family, while “DDPM” identifies a specific framework within that family.
Students sometimes make the opposite mistake and describe DDPM as the neural network doing the denoising. The denoiser, typically a convolutional neural network in the case study's explanation, is a component used within the process. DDPM describes the probabilistic noising and reverse-generation framework, not merely the CNN.
A concise exam distinction would be: “A diffusion model is the general class of models that generates data by reversing a noising process, whereas a DDPM is a probabilistic implementation that organizes this process into defined forward and reverse time steps.”
7 and 8: Generator vs discriminator
These are HL case study terms associated with generative adversarial networks.
| Generator | Discriminator |
|---|---|
| Produces synthetic candidate images from an input such as a multidimensional noise vector. | Estimates whether an input image is real training data or generated data. |
| Tries to create outputs convincing enough to fool the discriminator. | Tries to distinguish genuine examples from generator outputs. |
| Its output is the synthetic image. | Its output is a judgement, score, or probability. |
The generator does not search a database and copy the closest image. It learns a mapping from an input space, often called a latent space, to candidate images. Likewise, the discriminator does not generate corrections or repair the image; its feedback influences learning through the training objective.
The original Generative Adversarial Nets paper describes the interaction as a two-player minimax game. The generator improves at misleading the discriminator, while the discriminator improves at detecting generated samples.
For Visionary Studios, a GAN may produce sharp images efficiently after training, but achieving balanced adversarial training can be difficult. An answer that merely says “the generator generates and the discriminator discriminates” is too circular. State the input, output, and competing objective of each network.
9 and 10: Training instability vs mode collapse
| Training instability | Mode collapse |
|---|---|
| A broad condition in which learning fails to converge reliably or fluctuates unpredictably. | A specific GAN failure in which the generator produces too little variety. |
| Can arise when the generator and discriminator improve at incompatible rates. | Occurs when a narrow set of outputs repeatedly succeeds in fooling the discriminator. |
| May cause oscillation, poor images, vanishing gradients, or failed convergence. | Produces repetitive images that fail to represent the diversity of the training data. |
These terms are related, but they are not synonyms. Mode collapse is one possible outcome of unstable GAN training, while instability includes a wider range of optimization problems.
Suppose Visionary Studios needs 50 visually distinct characters for a campaign. A mode-collapsed generator might repeatedly produce similar faces, poses, or colour schemes even though each individual image appears realistic. The model has found a limited output pattern that fools the discriminator but has not captured the full distribution of possible designs.
This distinction supports evaluation. Sharpness alone would not make the GAN suitable if low diversity prevents the company from producing varied campaign assets. HL answers should therefore weigh image quality against diversity, training reliability, time, and the expertise required to balance both networks.
Other mistakes that weaken case study answers
Treating data curation and bias mitigation as identical
Dataset curation is the process of selecting, checking, organizing, documenting, and potentially removing training data. Bias mitigation is the broader effort to identify and reduce unfair or systematically distorted outcomes. Curation can support mitigation, but mitigation may also involve testing, reweighting, resampling, model adjustments, output filtering, and human review.
The NIST Generative AI Profile recommends examining data provenance, representativeness, intellectual property, and bias throughout development and evaluation. This matters for Visionary Studios because a large dataset is not automatically diverse, lawful, or suitable for commercial use.
Confusing realism with character consistency
Photorealism concerns whether an image appears visually convincing. Character consistency concerns whether identifying features remain stable across multiple images. A model could generate one highly realistic character but change that character's face, clothing, or proportions in the next output.
Writing definitions without scenario application
After defining a term, add a consequence for Visionary Studios. For example: “Conditional generation uses an input such as a sketch to constrain the output, allowing the studio to preserve a client's required composition.” This converts recalled knowledge into applied case study analysis.
A reliable method for learning IB CS key definitions
Use a three-part structure for every term:
- Category: What kind of concept is it?
- Mechanism: What does it do, and how?
- Contrast or application: How is it different from a related term, or why does it matter in the scenario?
For example: “A discriminator is a neural network within a GAN that estimates whether an image comes from the real training dataset or the generator. Unlike the generator, it judges images rather than creating them. Its feedback helps train the generator to produce more convincing designs.”
Revision should combine recall with application. The RevisionDojo A4 Machine Learning hub provides a useful starting point, while IB Computer Science study notes can reinforce the underlying syllabus concepts.
Use A4 Machine Learning flashcards for definition recall, then move to the A4 Machine Learning Questionbank to practise applying terminology. The broader Computer Science first-exam-2027 resource hub can be used to build timed practice, with Jojo AI helping identify vague definitions and missing scenario links.
Conclusion
The ten most important case study terminology mistakes arise when students merge concepts that perform different roles. Conditional generation requires a guiding input, noise injection and denoising are opposite processes, DDPM is a specific diffusion framework, and a GAN's generator and discriminator have competing functions. Training instability is a broad problem, while mode collapse is a particular loss of output diversity.
Learn each definition with its mechanism, contrast, and relevance to Visionary Studios. RevisionDojo's Study Notes and Flashcards can strengthen recall, while the Questionbank and Jojo AI are most useful for testing whether you can apply these distinctions in complete Paper 1 responses.
Sources and referenced URLs
- IB Computer Science curriculum updates
- Publicly accessible copy of the 2027 IB Computer Science case study booklet
- Denoising Diffusion Probabilistic Models research paper
- Generative Adversarial Nets research paper
- NIST Artificial Intelligence Risk Management Framework: Generative AI Profile
- RevisionDojo A4 Machine Learning hub
- RevisionDojo IB Computer Science study notes
- RevisionDojo A4 Machine Learning flashcards
- RevisionDojo A4 Machine Learning Questionbank
- RevisionDojo Computer Science first-exam-2027 resources
