The 2027 IB Computer Science case study, Generative AI for image creation: A diffuse vision, places you inside Visionary Studios, a company evaluating generative AI for advertising, concept art and digital media. Your final review should focus on four things: how the models work, how Visionary Studios should compare them, the four stated challenges, and how to turn technical knowledge into a justified recommendation.
For both SL and HL, the case study is assessed in Paper 1, Section B. SL students prepare the first two challenges, while HL students prepare all four. This case study last-minute revision sheet condenses the essential models, terminology, trade-offs and exam method into a format you can review immediately before the examination.
Paper 1 facts you must know
Under the syllabus first assessed in 2027, the case study is no longer part of a separate HL Paper 3. It appears in Paper 1 for both SL and HL, as confirmed by the IB Computer Science curriculum update.
RequirementSLHLPaper 1 duration1 hour 15 minutes2 hoursPaper 1 weighting35%40%Total Paper 1 marks5080Case-study marks1224Challenges studied24Case-study teaching time15 hours30 hours
All questions are compulsory. Section A covers Theme A concepts, while Section B uses the pre-seen case study. Do not revise the case study as an isolated story: connect it particularly to A4 Machine learning, computer hardware, computational resources, data quality and ethics.
The one-page exam day summary sheet
Scenario and decision
Visionary Studios wants to streamline the creation of high-quality images while protecting image quality, brand identity and public trust. It is evaluating diffusion models, GANs and hybrid models, although GANs and hybrids are HL-only areas in this case study.
Every recommendation should answer the same underlying question: Which approach best meets the company’s creative requirements at an acceptable computational, operational, ethical and legal cost? Avoid claiming that one model is universally best. The answer depends on the intended use, available infrastructure and importance of factors such as photorealism, speed, control and consistency.
Image-generation methods
MethodEssential meaningLikely Visionary Studios useText-to-image generationProduces an image from a written prompt.Rapid concept art and advertising prototypes.Conditional generationGenerates an image using an input such as a label, sketch, prompt or segmentation map.Controlled branded assets that must meet a brief.Unconditional generationGenerates from learned patterns without a specific conditioning input.Open-ended experimentation, but with less user control.Image-to-image translationConverts an input image into a related output, often changing style or visual domain.Turning sketches into polished concepts or applying a campaign style.
A segmentation map divides an image into labelled regions, such as person, sky and background. It can condition a model so that generated objects appear in specified locations. This gives designers more structural control than an unrestricted text prompt.
Diffusion model process: SL and HL
A diffusion model learns to reverse a gradual noising process. During generation, it begins with random noise and repeatedly predicts and removes noise until a coherent image appears.
-
Noise injection corrupts training images across multiple steps.
-
A neural denoiser, commonly a convolutional neural network (CNN), learns to predict noise or recover image structure.
-
Generation starts from random noise rather than a stored image.
-
Iterative denoising progressively creates recognisable features.
-
Conditioning information, such as a text embedding, steers the result.
A denoising diffusion probabilistic model (DDPM) represents this forward and reverse process probabilistically. The original DDPM research paper provides the technical foundation, while Figure 1 in the 2027 case-study booklet shows the transition from noise to a generated image.
Strengths: high-quality and often photorealistic outputs, good coverage of diverse image patterns, and strong compatibility with conditioning. Limitations: repeated denoising requires time, memory and processing power, making large-scale or real-time generation difficult. Possible optimizations include fewer sampling steps, latent-space processing, efficient schedulers, quantization, lower numerical precision and stronger GPU infrastructure, but each can create a quality, cost or compatibility trade-off.
GANs: HL only
A generative adversarial network (GAN) contains two competing neural networks:
-
The generator converts a D-dimensional noise vector into a synthetic image.
-
The discriminator compares real training images with generated images and predicts whether each is real or synthetic.
-
Through this adversarial dynamic, the generator improves at deceiving the discriminator while the discriminator improves at detecting generated outputs.
GANs can produce sharp images quickly after training, but training is sensitive. If the discriminator becomes too strong, the generator may receive weak learning signals; if the generator dominates, image quality may not improve reliably. Mode collapse occurs when the generator repeatedly produces a narrow set of similar outputs rather than representing the training dataset’s diversity.
Possible responses include adjusting the relative learning rates, balancing training updates, improving dataset diversity, applying regularization, adding controlled noise or using alternative loss functions. These measures can improve stability, but they require experimentation and do not guarantee convergence. Google’s guide to common GAN problems explains the relationship between unstable training, vanishing gradients and mode collapse.
Hybrid models: HL only
A hybrid model combines architectures so that one model compensates for another’s limitations. A strong answer explains the function of each component instead of merely stating that combining models improves performance.
ComponentCore strengthMain limitationPossible hybrid roleVAEStable training and structured latent space.Outputs may appear blurred.Encodes images into a compact representation.GANSharp outputs and fast inference.Unstable training and mode collapse.Improves perceptual sharpness.Flow-based modelExact likelihood and reversible mapping.High memory and architectural cost.Provides controllable, traceable latent transformations.Diffusion modelHigh quality, diversity and conditioning.Slow iterative sampling.Refines or generates detailed final images.
A latent space is a compressed mathematical representation in which meaningful image features are encoded as vectors. A VAE could compress an image into latent space, diffusion could generate or refine the latent representation, and a decoder could reconstruct the final image. The resulting system may improve efficiency and control, but integration complexity, training cost and maintenance also increase.
The four challenges and defensible responses
Case-study challengeLevelHigh-value response directionManage iterative denoising and computational demands while producing photorealistic images.SL and HLUse efficient sampling, suitable GPUs, latent diffusion, caching or reduced precision; test any speed-quality trade-off.Navigate large-dataset ethics, especially intellectual property and bias.SL and HLCurate licensed data, document provenance, remove unsuitable content, audit outputs, involve diverse reviewers and disclose AI use.Balance the GAN generator and discriminator to produce realistic images.HL onlyMonitor both losses, tune learning rates and update frequency, use regularization, test diversity and detect mode collapse.Evaluate how a hybrid could balance VAEs, GANs, flow-based models and diffusion.HL onlyAssign each component a justified role, then compare quality and control gains against complexity, compute and stability.
For the first challenge, optimization must be connected to Visionary Studios. Faster generation supports tight advertising deadlines and rapid client revisions, but excessive reduction in denoising steps may reduce detail or prompt adherence.
For the ethics challenge, dataset curation means intentionally selecting, documenting, cleaning and reviewing training data. Visionary Studios should verify licensing and consent, record data provenance, remove duplicates or harmful material, test demographic representation and provide a process for complaints or removal requests. The NIST Generative AI Profile supports lifecycle controls involving intellectual property, provenance, bias measurement, transparency and human oversight.
Five evaluation criteria to apply in every answer
The case study identifies five considerations. Use them as an instant planning framework for any extended response.
-
Output quality: resolution, photorealism, detail, stylistic coherence and alignment with the brief.
-
Computational efficiency: training and inference time, memory, processing hardware, energy use and cost.
-
Training stability: reliable convergence rather than repetitive, low-quality or unpredictable outputs.
-
Flexibility and scalability: adaptation to different campaigns and increasing production demand.
-
Consistency: stable characters, motifs, composition and brand style across multiple images.
Character consistency means preserving the recognisable appearance of a recurring character across poses, settings and outputs. An embedding-based approach represents a character or style as a learned vector that can condition future generations. This may improve continuity, but the studio still needs systematic human review because an embedding cannot guarantee perfect identity or style adherence.
Essential terminology in rapid-review form
TermExam-ready definitionBias mitigationMeasures used to detect, reduce and monitor unfair patterns in data or outputs.CNNA neural network that learns spatial features using convolutional filters.DenoisingRemoving predicted noise to recover or generate coherent image structure.DDPMA probabilistic model that learns to reverse a staged noising process.EmbeddingA vector representation encoding semantic, stylistic or identity-related features.Image generationCreation of new visual outputs from learned data patterns.Style adherenceThe extent to which outputs maintain the requested visual style.Training stabilityThe ability of training to converge reliably to useful performance.VAEAn encoder-decoder model that learns a probabilistic latent representation.Flow-based modelA generative model using invertible transformations between data and latent distributions.
Do not define a term and stop. Apply it immediately: “Mode collapse would be damaging because Visionary Studios needs diverse concepts for different clients, not repeated variations of one image.”
Last-minute answer method
For a short response, match your depth to the command term. Identify gives the point, outline gives the point plus a brief feature, and describe provides developed characteristics. Use the terminology from the case study rather than vague phrases such as “the AI becomes better.”
For an extended response, use this compact structure:
-
Judgement: state the most suitable position for Visionary Studios.
-
Technical explanation: explain how the relevant model works.
-
Benefits: connect advantages to quality, deadlines, control or scalability.
-
Limitations: discuss compute, instability, bias, IP or consistency.
-
Mitigation: propose realistic technical and governance controls.
-
Conclusion: weigh the trade-offs and give a conditional recommendation.
Avoid model dumping, unsupported claims that one architecture is always superior, and generic ethics paragraphs. Every paragraph should contain a case-study link such as advertising deadlines, recurring characters, client trust, branded assets or existing infrastructure.
For final practice, use the Computer Science First Exam 2027 resources to check concepts, then answer selected questions from the new-syllabus Computer Science Questionbank. A timed Computer Science predicted paper is more useful than rereading notes if you have already learned the terminology.
Conclusion
Your exam day summary sheet should reduce A diffuse vision to a decision framework: understand each model, apply the five evaluation criteria, address the challenges for your level, and reach a justified recommendation for Visionary Studios. SL students need diffusion, image-generation methods and the first two challenges; HL students additionally need GANs, hybrid architectures, VAEs, flow-based models and the final two challenges.
Use your remaining time to retrieve definitions from memory and write one balanced response rather than passively rereading the booklet. RevisionDojo’s IB Computer Science resource hub and Questionbank with Jojo AI feedback can support a final targeted check, followed by a short timed Paper 1 practice.