The May and November 2027 IB Computer Science case study, Generative AI for image creation: a diffuse vision, asks students to understand how diffusion models create images and how generative systems can be evaluated. Higher Level students must also study generative adversarial networks (GANs) and hybrid models.
The fastest mental map is this: a diffusion model repeatedly turns noise into an image, a GAN trains a generator against a discriminator, and a hybrid model combines techniques to balance quality, speed, control and stability. This diffusion models, GANs and hybrid overview establishes that mental map before introducing the terminology and comparisons needed for Paper 1.
What the 2027 IB case study requires
The case study places you in the context of Visionary Studios, a creative design company evaluating generative AI for advertising, concept art and digital media. It considers image quality, computational demands, control, training data and ethical or legal risks.
The level distinction is important:
ContentSLHLDiffusion models and iterative denoisingRequiredRequiredDDPMs and the neural denoiserRequiredRequiredConditional, unconditional and text-to-image generationRequiredRequiredModel evaluation, intellectual property and biasRequiredRequiredGAN generator, discriminator and mode collapseNot required by this case studyRequiredHybrid models, including their component approachesNot required by this case studyRequired
The case study is assessed in Paper 1 for both levels. The IB confirms that the former Paper 3 has been removed under the course first assessed in May 2027, with the case study moved into Paper 1. Although GANs and hybrids are outside the SL case-study scope, an SL student may still find the broad comparison useful.
Students should obtain the official case-study booklet through their teacher or IB school resources. A publicly accessible copy of the 2027 case-study booklet is also useful for checking its wording and line references.
The five-minute mental map
Imagine three ways of asking a machine to create a photograph of a futuristic city.
-
Diffusion model: Begin with television-like static and repeatedly clean it until a city appears.
-
GAN: Employ a forger to create city images and a detective to identify the fakes. Both improve through competition.
-
Hybrid model: Give different parts of the task to different specialists, perhaps compressing the image first, generating it through diffusion, and using adversarial training to sharpen details.
These are not three interchangeable labels. They describe different architectures and training processes, which leads to different strengths and limitations.
Model familyCore ideaMain strengthMain limitationDiffusionIteratively remove noiseHigh quality, diversity and controlRepeated denoising requires computationGANGenerator competes with discriminatorSharp images and potentially fast generationSensitive training and mode collapseHybridCombine complementary approachesCan balance quality, speed and controlGreater design and training complexity
Diffusion models: noise becomes structure
A diffusion model is a generative model that learns to reverse a gradual noising process. NIST describes diffusion models as having a forward process, a reverse process and a sampling procedure.
Training and generation are related but different
During training, a clean training image is progressively corrupted by adding noise. A neural denoiser, typically built using a convolutional neural network, learns to predict or remove the noise at different stages.
During generation, the system does not need to retrieve a particular training image. It begins with random noise and repeatedly applies its learned reverse process until a new coherent image appears. Text or another condition can guide what the denoising process produces.
A useful sequence is:
-
Take clean images from the training dataset.
-
Add controlled noise over a sequence of steps.
-
Train the neural network to estimate how the noise can be reversed.
-
Begin generation from fresh random noise.
-
Apply iterative denoising until an image is formed.
Modern diffusion systems build on the denoising diffusion probabilistic model (DDPM) framework. The foundational DDPM research formalized image generation as a learned reverse process corresponding to a predefined forward corruption process.
Why diffusion models are useful
Diffusion models can produce detailed, varied images and can support strong conditioning. Conditioning means guiding generation using information such as a text prompt, class label, reference image or edge map.
Their main disadvantage follows directly from their design. If generation requires many sequential denoising steps, the model must repeatedly run a neural network. That can increase latency, energy use and hardware requirements, although newer sampling and latent-space techniques can reduce the number or cost of these operations.
For an exam answer, connect cause and consequence: iterative denoising increases computational demand because each step requires another model evaluation.
GANs: a generator competes with a discriminator
This section is HL case-study content. A generative adversarial network contains two neural networks trained through an adversarial process:
-
The generator converts a random latent or noise vector into a candidate image.
-
The discriminator receives real training images and generated images, then attempts to distinguish between them.
The generator improves by learning to produce outputs that fool the discriminator. The discriminator improves by detecting weaknesses in generated outputs. The original GAN paper describes this as a two-player minimax game.
The simple analogy of a forger and detective is helpful, but an exam answer should explain the information flow. Random input enters the generator, the generated image is assessed by the discriminator, and the discriminator's judgement supplies a learning signal that updates the generator.
Instability and mode collapse
GAN training can become unstable if one network becomes much more capable than the other. If the discriminator rejects almost every generated sample too confidently, the generator may receive an unhelpful learning signal. If the generator exploits one narrow weakness in the discriminator, it may repeatedly produce similar outputs.
This second problem is called mode collapse. It occurs when many different input vectors map to the same or a limited range of outputs, so the generated dataset lacks the diversity of the real distribution. For Visionary Studios, a GAN that repeatedly creates similar faces, layouts or colour schemes would be commercially restrictive even if each individual image looked realistic.
Possible mitigation strategies include adjusting the relative training rates, changing loss functions, regularization, increasing dataset diversity and using alternative GAN architectures. In Paper 1, the important point is not to list techniques without explanation, but to show how they help maintain a productive balance between generator and discriminator.
Hybrid models: combining complementary strengths
This section is also HL case-study content. A hybrid generative model combines two or more model families or training approaches so that one component compensates for another's limitation. It is therefore a design category, not one fixed architecture.
Relevant components include:
-
Variational autoencoders (VAEs): Encode images into a compact latent space and decode latent representations back into images. They offer structured representations and efficient manipulation, although standalone outputs may lose fine detail.
-
Flow-based models: Learn invertible transformations between a simple probability distribution and the data distribution. They can support exact likelihood calculation, but invertibility constraints can make architecture design demanding.
-
GANs: Contribute adversarial training that can encourage sharp, realistic detail and fast generation.
-
Diffusion models: Contribute stable denoising, diversity and flexible conditioning, but may require several sampling steps.
A common conceptual hybrid is a latent diffusion model. A VAE-like encoder first compresses an image into a smaller latent representation, diffusion operates in that reduced space, and a decoder reconstructs the final image. This reduces computation compared with performing every denoising operation directly on full-resolution pixels.
Another hybrid can incorporate an adversarial objective to encourage sharp output or reduce the number of generation steps. However, adding components does not automatically create a better system. It can introduce more hyperparameters, interacting loss functions, debugging difficulty and training cost.
How to compare models in the case-study scenario
Avoid claiming that one architecture is always best. Visionary Studios should select a model according to its intended task and constraints.
Decision factorQuestion to askImage qualityAre details coherent and suitable for professional use?DiversityDoes the system cover varied styles, subjects and layouts?SpeedHow long does one image or batch take to generate?Computational costWhat hardware, memory and energy are required?ControllabilityCan prompts or conditions reliably direct the output?Training stabilityCan the model be trained consistently?Data governanceAre images licensed, representative and documented?Practical complexityCan the company maintain and monitor the system?
Suppose Visionary Studios needs rapid generation of many similar product mock-ups. A well-trained GAN might offer fast inference, but limited diversity or unstable retraining could be concerns. If the company needs highly controllable concept art across varied prompts, diffusion may be preferable despite slower generation. A hybrid may offer a better compromise, provided the company can manage its additional complexity.
Common mistakes in Paper 1 answers
-
Confusing training with generation: Noise is added to clean images in the forward training process; new generation starts from noise and follows the learned reverse process.
-
Swapping GAN roles: The generator creates candidates, while the discriminator classifies real and generated inputs.
-
Defining mode collapse as low image quality: It primarily concerns limited output diversity, even though the repeated outputs may look sharp.
-
Calling every modern model a hybrid: A hybrid must meaningfully combine model types, representations or objectives.
-
Giving absolute judgements: Model performance depends on architecture, dataset, implementation, hardware and task.
-
Ignoring the scenario: Apply every comparison to Visionary Studios rather than writing a detached technical description.
-
Forgetting ethics: Copyright, consent, bias, provenance and misleading synthetic media remain relevant regardless of architecture.
Use command terms carefully. Describe requires characteristics or a process, explain requires reasons or mechanisms, and evaluate requires balanced arguments followed by a supported judgement.
A practical revision method
Start by memorizing one precise sentence for each model, then practise expanding it into a process, strength, limitation and scenario-based recommendation. Draw the information flow from memory: noise to repeated denoising for diffusion, and noise to generator to discriminator feedback for a GAN.
The RevisionDojo Computer Science hub for the 2027 course can support this progression with notes, lessons and flashcards. Follow conceptual review with IB Computer Science practice resources, then use Computer Science predicted papers for timed application. Jojo AI can help turn definitions into retrieval questions, but compare its feedback against the official case-study wording and your teacher's guidance.
Conclusion
Diffusion models generate images through iterative denoising, GANs train a generator and discriminator in competition, and hybrid models combine complementary techniques. For HL, the crucial additions are adversarial training, mode collapse, latent representations and the reasons for combining VAEs, flows, GANs and diffusion.
Strong Paper 1 answers connect architecture to consequences: quality, diversity, speed, computation, control, stability and ethical risk. RevisionDojo Study Notes, Flashcards, Questionbank practice and Jojo AI can help you move from this quick guide to accurate, scenario-based exam responses.
