Case Study 2027: Generative AI for Image Creation

"A diffuse vision". For Paper 1, May and November 2027. Everything in the booklet, explained, plus model answers.

⬇ DOWNLOAD THE OFFICIAL CASE STUDY BOOKLET (PDF)8 pages

How to use this page

SL + HL sections are for everyone. HL ONLY sections (GANs, hybrid models, VAEs, flow-based models) are marked in a dashed gold box.
Line numbers like (L35) point to the booklet, so you can find the exact wording.
The booklet is used in Paper 1 at both SL and HL. Questions mix the case study with the syllabus, especially A4 Machine learning.
Read the whole page once, learn the glossary, then do the practice questions without looking.

Contents

The scenario in one minute
Background you must know first
Image generation techniques
Diffusion models (the core of the case study)
GANs HL
Hybrid models, VAEs, flow-based models HL
All models compared
Evaluating generative AI models
Ethical and legal considerations
The challenges, answered
Glossary: every term in the booklet
Practice questions with model answers
How to get full marks

1. The Scenario in One Minute

Visionary Studios is a creative design company (L10). It wants to use generative AI to produce high quality images faster for:

advertising campaigns
concept art
digital media and branded assets

It is comparing three families of models: diffusion models, GANs (HL) and hybrid models (HL). It will judge them on output quality, computational efficiency, training stability, flexibility and scalability, and consistency. It must also deal with copyright, bias and AI disclosure.

Ninja Note: every answer should be about Visionary Studios. Generic answers lose marks. Name its real needs: brand identity, recurring characters, advertising deadlines, client trust, its existing infrastructure.

2. Background You Must Know First

Concept	What it means here
Generative AI	AI that creates new content (images, text, audio) similar to its training data, rather than only classifying existing data.
Generative vs discriminative model	A generative model learns what the data looks like so it can produce new examples. A discriminative model learns to tell classes apart (cat or dog?). A GAN uses one of each.
Neural network	Layers of connected nodes with weights that are adjusted during training to reduce error (see A4.1).
Training vs inference	Training: adjusting weights using a dataset (slow, done once). Inference / generation: using the trained model to make new images (done every time a designer asks for one).
Training dataset	The images (often with captions) the model learns from. The model can only generate what its data taught it (L41).
Noise	Random pixel values, like TV static. Generative models use noise as a random starting point, which is why every generated image is different.
Image as data	An image is a grid of pixels, each with RGB values (see A1.2). A 512×512 colour image is 786,432 numbers.
GPU / TPU	Hardware for massive parallel matrix calculations. Generative models are trained and run on them (see A4.1.3).

Convolutional neural network (CNN)

A CNN is a neural network designed for image data. It slides small filters (kernels) across the image to detect features. Early layers find edges and textures, deeper layers find shapes and whole objects. Because the same filter is reused across the image, it handles large images with far fewer weights than a normal network.

In a diffusion model the CNN is the neural denoiser (L38): it looks at a noisy image and predicts the noise in it. The usual CNN shape for this is a U-Net: it shrinks the image down to capture the overall layout, then expands it back up to full size to restore fine detail.

Embeddings

An embedding is a list of numbers (a vector) that represents something such as a word, a prompt, a character or a style. Similar things get similar vectors. Text-to-image models turn the prompt into an embedding, and embedding-based approaches keep a character looking the same in every image (L92).

3. Image Generation Techniques SL + HL

The booklet sorts models by what you give them as input.

Technique	Input	How it works	Use at Visionary Studios
Text-to-image generation (L19–22)	A text prompt	The prompt is turned into an embedding by a text encoder. The embedding guides each denoising step so the image matches the description. Examples: DALL·E, Stable Diffusion.	Concept art and adverts with a specific theme or style. Fast exploration of ideas.
Conditional image generation (L23–29)	A condition: class label, sketch, segmentation map, or text	The condition is fed into the model with the noise, so the output is constrained to match it.	Outputs that must match a brief, layout or brand style.
• Class-conditional	A category label	Generates an image of a chosen class (e.g. "animal", "landscape").	Stock style images of a set category.
• Image-to-image translation	An image	Transforms one type of image into another: sketch → realistic render, black and white → colour.	Turning artists' rough sketches into finished renders, keeping their composition.
Unconditional image generation (L30–34)	Nothing except random noise	Generates whatever the training data taught it, with no control over the subject.	Exploratory creativity, abstract art, novel visuals, synthetic datasets for training other AI.

Segmentation map

A segmentation map is an image where every pixel is labelled with the class of object it belongs to, shown as flat colour regions (e.g. blue = sky, green = grass, grey = building). Used as a condition, it tells the model exactly where each thing goes, so a designer controls the layout and the AI fills in realistic detail.

Ninja Notes:

Conditional = control. Unconditional = creativity / variety. That trade off is the heart of any "compare" question.
Text-to-image is a type of conditional generation (the condition is the text). The booklet lists it separately, but say this in an answer and it shows understanding.
Diffusion is the method. Text-to-image, conditional and unconditional describe the input. A diffusion model can be any of the three.

4. Diffusion Models SL + HL

"Diffusion models generate images by leveraging an iterative denoising process, beginning with noise injection... which is refined over multiple steps into coherent outputs." (L36–38)

The big idea

Destroying an image with noise is easy. A diffusion model learns to reverse that destruction, one small step at a time. Once it can do that, it can start from pure noise and "undo" it into a brand new image that never existed.

Step by step

Stage	What happens
1. Forward process (training)	Take a real training image. Add a small amount of Gaussian (random) noise. Repeat for many steps (often around 1000) until it is pure noise. This follows a fixed noise schedule, so no learning is needed here.
2. Learning to denoise (training)	Show the CNN a noisy image and the step number. It predicts the noise that was added. Compare with the real noise, measure the error, adjust the weights. Repeat millions of times across the dataset.
3. Noise injection (generation)	Start with an image of pure random noise (L37). A different random start gives a different image.
4. Iterative denoising (generation)	The trained CNN predicts the noise, a little is removed, and the slightly cleaner image goes back in. Repeat for many steps. Structure appears first (layout, shapes), then fine detail.
5. Output	A coherent, photorealistic image. What it shows depends on the training data (trained on animals → cats and dogs; trained on cities → urban scenes) (L41–43) and on any condition such as a prompt.

Try it: the forward and reverse process

Drag the slider to add noise step by step (forward process). Press Denoise to watch the reverse direction. This is an illustration: a real model has to predict the noise, it does not know the original.

Step 0 / 100

Image kept: 100%

Noise: 0%

DDPM

The denoising diffusion probabilistic model (DDPM) is the framework modern diffusion models are based on (L46–47). It formally defines the two processes: gradually adding noise during training (forward) and gradually removing it during generation (reverse). "Probabilistic" because noise is random, so the model works with probabilities, and every run can give a different image.

Text-to-image with diffusion

Why diffusion models are computationally expensive

Iterative: one image needs many denoising steps (tens to hundreds at generation, often around 1000 in the original DDPM).
Each step is a full pass through a large CNN with millions or billions of weights, over every pixel.
High resolution multiplies the work: twice the width and height is four times the pixels.
Training needs huge datasets, many GPU hours, lots of memory (VRAM) and energy.
Compared with a GAN, which makes an image in one forward pass, diffusion is much slower to generate.

How to optimise the process (L44–45, L109–110)

Technique	How it helps	Trade off
Fewer sampling steps (faster samplers, e.g. DDIM)	Skip steps: 20 to 50 instead of 1000.	Too few steps lowers quality.
Latent diffusion	Run denoising on a compressed version of the image (latent space), not every pixel. Stable Diffusion does this. Far less computation.	Needs an encoder/decoder; tiny details can be lost.
Better hardware	GPUs/TPUs run the CNN's matrix maths in parallel.	Expensive to buy.
Cloud computing	Rent GPUs on demand, scale up for big campaigns (see A1.1).	Ongoing cost, data leaves the company.
Use a pre-trained model + fine-tune (transfer learning)	No need to train from scratch. Fine-tune on the studio's own style.	Inherits the original dataset's copyright and bias issues.
Distillation	Train a smaller, faster "student" model to copy a big model in fewer steps.	Extra training, slight quality loss.
Lower precision / smaller models	Use 16 bit numbers instead of 32 bit: less memory, faster.	Possible small accuracy loss.
Generate small, then upscale	Draft quickly at low resolution, upscale only the chosen images.	Upscaling can add artefacts.
Batching and caching	Generate several images in one run; reuse prompt embeddings.	Needs more memory per run.

Strengths and weaknesses

Strengths	Weaknesses
Excellent photorealistic, high quality output (L43–44)	Slow generation: many iterative steps
Stable training: one network, simple goal (predict the noise)	High computational and energy cost
High diversity: covers the whole dataset (no mode collapse)	Needs huge training datasets (copyright and bias risk)
Easy to condition on text, sketches, segmentation maps	Random start makes exact consistency harder without extra techniques

HL ONLY (L48–59)

5. Generative Adversarial Networks (GANs)

A GAN is two neural networks trained against each other:

Network	Input	Output	Goal
Generator	A D-dimensional noise vector	A fake (synthetic) image	Fool the discriminator into labelling its images "real"
Discriminator	Real images from the dataset and fake images from the generator	A predicted label: real or fake (a probability)	Correctly tell real from fake

Think of a forger (generator) and a detective (discriminator). Each improves because the other does.

Training loop (one round)

Generator turns random noise vectors into a batch of fake images.
Discriminator is shown real and fake images and labels each one. Its weights are updated to get better at spotting fakes.
Generator's weights are updated using the discriminator's verdict, so its next fakes are more convincing.
Repeat thousands of times. Ideally they reach a balance where fakes are so good the discriminator can only guess (50/50).

Key terms

D-dimensional noise vector: a list of D random numbers (e.g. 100) that is the generator's input. Each different vector gives a different image. The vector space it comes from is the generator's latent space.
Adversarial dynamic: the competition between the two networks. It is "key to a GAN's ability to produce sharp, high-quality images" (L55), because the generator is pushed until its images are indistinguishable from real ones.

Problems with GAN training (L56–59)

Problem	What happens
Training instability	The two networks keep overpowering each other, so quality swings up and down and may never converge. Very sensitive to settings like learning rate.
Discriminator too strong	It rejects every fake with total confidence, so the generator gets almost no useful feedback and stops improving (vanishing gradients).
Generator too strong	It fools a weak discriminator with poor images, so quality stays low.
Mode collapse	The generator finds a few outputs that fool the discriminator and only produces those, so outputs lack variety instead of covering the full diversity of the dataset. E.g. asked for faces, it makes the same few faces again and again.

Mitigation: balancing generator and discriminator (L113–115)

Tune learning rates separately, or update one network more often than the other, so neither dominates.
Improved loss functions (e.g. Wasserstein loss) that give the generator useful feedback even when the discriminator is winning.
Regularising the discriminator (e.g. spectral normalisation, label smoothing, adding noise to its inputs) so it cannot become overconfident.
Minibatch discrimination / feature matching: the discriminator compares groups of images, so a generator producing identical images is caught. Directly fights mode collapse.
Progressive growing: start training at low resolution and add layers gradually, which makes training more stable.
Monitoring: regularly check output variety and quality, save checkpoints, roll back if collapse starts.

GAN strengths	GAN weaknesses
Sharp, high quality images	Unstable, hard to train
Very fast generation: one forward pass, no iteration	Mode collapse: low diversity
Good for real-time uses and upscaling	Harder to condition on detailed text prompts

HL ONLY (L60–75). You need a broad understanding of VAEs and flow-based models, and a deep understanding of hybrid models.

6. Hybrid Models, VAEs and Flow-based Models

A hybrid model combines two or more types of generative AI so the strengths of one cover the weaknesses of another. This gives higher quality, better control over output attributes and more flexibility for different creative needs (L61–63).

Variational autoencoders (VAEs)

Encoder compresses an image into a small set of numbers in the latent space: "a compressed and organized version of the original data, where similar images are placed closer together" (L65–67).
Decoder turns a point in latent space back into an image.
Because the space is organised, you can move through it: blend two images, change one attribute, or pick a random point to generate something new.
Strengths: stable training, fast, a smooth latent space ideal for flexibility, experimentation and artistic exploration (L68–69).
Weakness: outputs tend to be blurry and less detailed than GANs or diffusion.

Flow-based models

A chain of invertible (reversible) transformations that maps random noise to a realistic image exactly (L70–71).
Because every step can be run backwards without loss, you can go image → noise as well as noise → image, and trace exactly how each image was formed (L71–72).
Strengths: precision, reproducibility, transparency (L72–73): the same input always gives the same image; you can explain where it came from.
Weaknesses: every layer must be reversible, which limits the architecture, so models are large and memory hungry, and image quality is usually below diffusion and GANs.

What a hybrid combines (L74–75, L116–117)

Model	Brings	Weakness the others cover
Diffusion	Photorealism, diversity, stable training	Slow, expensive
GAN	Sharp detail, very fast generation	Unstable, mode collapse
VAE	Compact, organised latent space, flexible editing	Blurry outputs
Flow-based	Exact, reversible, reproducible, transparent	Heavy, lower quality

Real examples of hybrids:

VAE + diffusion = latent diffusion (how Stable Diffusion works): the VAE encoder compresses the image into latent space, diffusion runs there (much cheaper), and the VAE decoder rebuilds the full image. Solves the computational demand challenge.
VAE + GAN: a GAN discriminator is added to a VAE to make its blurry outputs sharp, while keeping the VAE's stable training and editable latent space.
Diffusion + GAN: a GAN style discriminator lets a diffusion model make good images in far fewer steps, or a GAN upscaler sharpens diffusion drafts quickly.
Flow-based component: adds reproducibility and traceability, useful for transparency and AI disclosure.

Downsides of hybrids: more complex to design, train and maintain; more components that can fail; may need more expertise; combined training can still be unstable.

7. All Models Compared

SL students only need the diffusion column. HL students should be able to reproduce this whole table.

Criterion	Diffusion	GAN (HL)	VAE (HL)	Flow-based (HL)	Hybrid (HL)
Output quality	Very high, photorealistic	High, sharp	Lower, blurry	Moderate	Highest potential
Generation speed	Slow (many steps)	Very fast (one pass)	Fast	Moderate	Depends on mix
Training stability	Stable	Unstable	Stable	Stable	Can be complex
Diversity	High	Risk of mode collapse	Good	Good	High
Control / editing	Good with conditioning	Harder	Excellent (latent space)	Exact mapping	Best
Transparency	Low	Low	Medium	High (reversible)	Depends on mix
Compute cost	High	Medium	Low	High memory	Higher design cost
Best for Visionary Studios	Final photorealistic advert images, concept art from prompts	Fast drafts, upscaling, real-time previews	Exploring styles, variations of a character	Reproducible, auditable brand assets	A full production pipeline

8. Evaluating Generative AI Models SL + HL (L76–95)

Factor	Question Visionary Studios asks	What affects it
Output quality	Are images high resolution, photorealistic, stylistically coherent and on brief?	Model type, size and quality of training data, number of denoising steps, resolution.
Computational efficiency	Can existing infrastructure train and deploy the model without big delays?	Number of steps, model size, GPUs/VRAM available, cloud vs on-site, latent vs pixel space.
Training stability	Does training reliably converge to good results?	GANs are unstable and can mode collapse; diffusion is stable but slower and more resource hungry (L83–85).
Flexibility and scalability	Can one model serve adverts, concept art and branded assets? Can it handle bigger future projects?	Conditioning options, ability to fine-tune, cloud scaling, cost per image.
Consistency	Does every image keep the brand identity?	Character consistency and style adherence (below).

Consistency in detail

Character consistency (L91–93): a recurring mascot or character must look the same in every image. Because generation starts from random noise, the same prompt can give a different face each time. Fixes:

Embedding-based approaches: learn an embedding (a vector) that represents the character from a few reference images, then include it in every generation as a condition. The model then "knows" that character.
Fine-tuning the model on images of the character.
Reference image conditioning or image-to-image from an approved design.
Fixing the random seed so the starting noise is the same.

Style adherence (L94–95): conditional models must produce images that match the style or theme in the prompt. Fixes:

Clear, detailed prompts and a studio prompt template.
A stronger guidance setting, so the model follows the condition more closely.
Style embeddings or fine-tuning on the brand's own artwork.
Human review before anything is published.

Quality can also be measured: human ratings by art directors, and automatic scores that compare generated images with real ones (e.g. FID, lower is better) or check how well an image matches its prompt.

9. Ethical and Legal Considerations SL + HL (L96–105)

Issue	The problem	What Visionary Studios should do
Dataset curation	Large datasets are often scraped from the internet and contain copyrighted art and photos. Using them may break intellectual property (IP) law, and the model can reproduce near copies of an artist's work or style.	Use licensed, public domain or own images; keep records of where data came from (provenance); respect artist opt-outs; filter out copyrighted and watermarked images; choose vendors that disclose their training data; legal review.
Bias and fairness	If the dataset over-represents some groups, outputs will too: e.g. "CEO" generating only one gender or skin tone, or stereotyped portrayals of cultures. This is exclusionary or inaccurate representation, harms people and damages the brand.	Audit datasets for balance; add diverse data or rebalance; test outputs across many prompts; diverse review team; human approval before publishing; feedback channel for complaints.
Transparency and AI disclosure	Clients and audiences may feel deceived if they do not know images are AI generated. Trust and reputation are at risk.	Clear policy on when AI is used; label AI content; tell clients in contracts; keep generation records; use watermarks or content credentials (metadata).

Other points worth adding in a "discuss" answer

Ownership: who owns an AI image? In many countries pure AI output may not be protected by copyright, so a client may not get exclusive rights.
Jobs: illustrators and photographers may lose work; AI can also free them for higher level creative tasks.
Misuse: realistic images can be used for deepfakes and misinformation (see A4.4).
Environment: training and running diffusion models uses a lot of energy.
Privacy: datasets can contain photos of real people without consent.

Bias mitigation = the steps taken to reduce bias: before training (curate and balance data), during training (reweight under-represented data), after training (test outputs, filter, human review).

10. The Challenges, Answered (L106–117)

These four bullet points are where the extended questions come from. Know an answer for each.

The problem: each image needs many denoising steps, each a full pass through a large CNN, so generating photorealistic, high resolution images is slow and needs powerful GPUs and lots of energy. Training is even more demanding.

Solutions: reduce the number of steps with faster samplers; use latent diffusion to denoise a compressed image; use a pre-trained model and fine-tune instead of training from scratch; run on GPUs/TPUs or scalable cloud computing; generate low resolution drafts and upscale only approved images; distil to a smaller model; batch requests.

Trade off: fewer steps or compression speed things up but can reduce quality and detail. The studio must balance speed (advertising deadlines, cost) against quality (photorealism, brand standards). A sensible plan: fast, low step drafts for concept exploration, full quality only for final assets.

IP: large datasets may include copyrighted images used without permission; the model may reproduce protected work or an artist's style, exposing the studio and its clients to legal action. Fix: curate datasets, use licensed or owned data, record provenance, respect opt-outs, check outputs for near copies.

Bias: unbalanced data leads to stereotyped or exclusionary images that can offend audiences and damage brands. Fix: audit and rebalance data, test outputs across many prompts and groups, diverse human review, clear guidelines.

Plus transparency: disclose AI use to clients and audiences to maintain trust.

Trade off: curated, licensed datasets are smaller and more expensive, which may lower quality or variety, but they reduce legal and reputational risk.

If the discriminator becomes too strong, it rejects everything and the generator receives no useful feedback, so learning stops. If the generator becomes too strong, it fools a weak discriminator with poor images. Imbalance also causes instability (quality oscillates) and mode collapse (the generator repeats a few images).

Solutions: tune each network's learning rate or update frequency; use improved loss functions (e.g. Wasserstein) that keep feedback useful; regularise the discriminator (spectral normalisation, label smoothing, input noise); minibatch discrimination to detect lack of variety; progressive growing; monitor diversity and use checkpoints.

For Visionary Studios: GANs give sharp images very quickly, useful for fast drafts or upscaling, but the training effort and risk of low variety mean they may be better as one part of a hybrid than as the main model.

Each model has a weakness another covers. A possible pipeline for Visionary Studios:

VAE encoder compresses images into an organised latent space (cheaper computation, easy editing and style exploration).
Diffusion runs in that latent space, conditioned on the prompt and on character/style embeddings (photorealism, diversity, stable training, style adherence).
GAN component sharpens detail or reduces the number of steps (speed, sharpness).
VAE decoder rebuilds the full resolution image.
Flow-based component where reproducibility or traceability is needed (transparency, AI disclosure, recreating an approved asset exactly).

Evaluation: gains in quality, control, speed and consistency; but higher design complexity, more expertise needed, more parts to maintain and test, and combined training can still be unstable. Worth it for a studio producing large volumes of branded content; probably overkill for occasional use, where a pre-trained latent diffusion model is enough.

11. Glossary: Every Term in the Booklet

These are the booklet's "additional terminology" (page 8). Expect "define" and "outline" questions on any of them.

SL + HL

Term	Definition
Bias mitigation	Techniques used to detect and reduce bias in training data and model outputs, such as balancing datasets, auditing and testing outputs, so results are fair and representative.
Character consistency	The ability of a model to generate the same recurring character or motif with a uniform appearance across multiple images.
Convolutional neural network (CNN)	A neural network designed for image data that uses filters (convolutions) to detect features such as edges, shapes and objects. Used as the denoiser in diffusion models.
Dataset curation	Carefully selecting, checking and cleaning training data so it is legal (no copyrighted material), relevant, high quality and free from bias.
Denoising	Removing noise from an image. In diffusion models, a neural network predicts and removes a little noise at each step to reveal a coherent image.
Denoising diffusion probabilistic model (DDPM)	The framework behind modern diffusion models that formalises gradually adding noise to images during training and gradually removing it during generation.
Diffusion model	A generative model that creates images by starting from random noise and iteratively denoising it over many steps into a coherent image, using patterns learned from training data.
Embedding-based approach	Representing a concept such as a character or style as an embedding (vector of numbers) that is reused as a condition, so generated images stay consistent.
Image generation	Using AI to create new images from learned patterns, with or without an input condition.
• Conditional	Image generation guided by a specific input (class label, sketch, segmentation map, text) so the output matches a predefined objective or style.
• Unconditional	Image generation with no predefined input, relying only on patterns learned from training data; used for exploration, abstract art and synthetic datasets.
Image-to-image translation	Conditional generation that transforms one type of image into another, e.g. sketch to realistic render, black and white to colour.
Noise injection	Introducing random noise as the starting point for generation (and adding noise to images during training).
Segmentation map	An image in which each pixel is labelled with the object class it belongs to, used as a condition to control the layout of a generated image.
Text-to-image generation	Generating an image from a text prompt, where the description guides the content and style of the output.
Training stability	How reliably a model's training converges to good results without failing, oscillating or producing repetitive outputs.

HL ONLY

Term	Definition
Adversarial dynamic	The competition between a GAN's generator and discriminator, where each improves by trying to beat the other.
D-dimensional noise vector	A vector of D random numbers used as the generator's input; each different vector produces a different image.
Generative adversarial network (GAN)	A generative model of two neural networks, a generator and a discriminator, trained together competitively to produce realistic images.
• Discriminator	The network that classifies images as real (from the dataset) or fake (from the generator).
• Generator	The network that turns a noise vector into a synthetic image, trying to fool the discriminator.
Hybrid model	A model that combines two or more generative approaches (e.g. VAE, GAN, flow, diffusion) to use their strengths and reduce their weaknesses.
• Flow-based model	A model that learns an exact, invertible mapping between random noise and images, so generation can be reversed without loss and traced.
• Variational autoencoder (VAE)	A model with an encoder that compresses images into a latent space and a decoder that generates images from points in it.
Latent space	A compressed, organised representation of data where similar images are placed close together.
Mode collapse	A GAN failure where the generator produces only a limited variety of outputs instead of the full diversity of the dataset.

12. Practice Questions with Model Answers

Written in the style of Paper 1. Mark points are shown in [ ]. Try each one before opening the answer.

SL + HL

Introducing random noise as the starting point from which an image is generated [1].

Any two: class label [1], sketch [1], segmentation map [1], text prompt [1], another image (image-to-image) [1].

The CNN acts as the neural denoiser, predicting the noise present in an image at each step [1], so it can be removed gradually to reconstruct image features using patterns learned in training [1].

Generation starts with an image of pure random noise (noise injection) [1]. A trained CNN denoiser predicts the noise in the image [1], and a small amount is removed, producing a slightly clearer image [1]. This is repeated over many iterations until a coherent image appears, whose content depends on the training data and any condition such as a prompt [1].

Conditional generation uses a specific input such as a sketch, label or prompt [1], so the output is controlled and matches a predefined objective or style [1]. Unconditional generation uses no input other than random noise [1], so outputs are based only on learned patterns, giving novel, unpredictable images useful for exploration or synthetic data [1].

The denoising process is iterative, so one image needs many steps [1]. Each step is a full pass through a large CNN with millions of parameters [1], applied to every pixel, so higher resolution images need much more computation [1]. Training also needs very large datasets and many GPU hours, using a lot of memory and energy [1].

Use latent diffusion [1], so denoising runs on a compressed representation rather than every pixel, needing far less computation [1]. Use cloud GPUs [1], which can be scaled up during busy advertising campaigns so many images are generated in parallel [1]. (Also accept: fewer sampling steps, pre-trained model, distillation, draft then upscale.)

The character's appearance is learned from reference images and stored as an embedding, a vector of numbers [1]. This embedding is included as a condition every time an image is generated [1], so the model reproduces the same features in each image, keeping recurring characters uniform for the brand [1].

The model learns patterns from its training data, so if some groups are under-represented [1], it will rarely generate them or show them stereotypically [1]. Adverts could then exclude or misrepresent parts of the audience [1], harming people and damaging the client's brand and the studio's reputation [1].

Datasets may contain copyrighted material, so curation avoids breaking intellectual property law and legal action [1]; it also removes biased or low quality data so outputs are fair and accurate [1].

For: diffusion models produce high quality, photorealistic images, matching the standard clients expect in adverts. Text-to-image and conditional generation let designers turn briefs, sketches or segmentation maps into images quickly, speeding up concept work. Training is stable and outputs are diverse, giving many options to choose from. Embedding-based approaches and conditioning help with character consistency and style adherence for brand identity.

Against: the iterative denoising process is slow and computationally expensive; the studio's existing infrastructure may not cope without GPUs or cloud costs. Random starting noise makes exact consistency difficult without extra techniques. Large training datasets raise copyright and bias risks, and clients may object if AI use is not disclosed.

Conclusion: yes, but managed: use a pre-trained latent diffusion model fine-tuned on licensed or owned brand imagery, run on scalable cloud GPUs, keep human review of every image for quality and bias, and disclose AI use. This gains speed and quality while controlling cost, legal and reputational risk.

Marks come from: balanced arguments for and against, each developed and linked to the studio (not generic), and a justified conclusion.

For full disclosure: maintains trust with clients and audiences; avoids accusations of deception; clients may need to know for legal reasons, since pure AI output may not be protected by copyright; supports ethical guidelines and future regulation.

Against / limits: some clients may value images less or demand lower prices; competitors may not disclose; when AI is only a small part of a heavily human edited image, labelling every asset may be impractical.

Judgement: disclose clearly to clients in every contract, and label to audiences where images are mostly AI generated or realistic enough to mislead; keep records of how each asset was made. Trust and legal safety outweigh the short term commercial risk.

HL ONLY

When a GAN's generator produces only a limited variety of outputs instead of capturing the full diversity of the training dataset [1].

The generator transforms a D-dimensional random noise vector into fake images [1]. The discriminator is shown both real images from the dataset and the fake images, and predicts which are real [1]. The discriminator is updated to classify more accurately [1], while the generator is updated using the discriminator's feedback to produce more convincing images; repeating this adversarial process improves both networks [1].

If the discriminator easily identifies every fake image [1], the generator receives very little useful feedback about how to improve [1], so its training stalls and image quality stops improving, making training unstable [1].

Minibatch discrimination: the discriminator looks at groups of images together [1], so a generator producing near identical images is easily detected and is pushed to create more variety [1]. (Also accept improved loss functions such as Wasserstein loss, or monitoring diversity and adjusting training.)

Both can produce high quality images [1]. GANs generate an image in a single forward pass, so they are much faster [1], while diffusion models need many iterative denoising steps and more computation [1]. Diffusion training is stable with diverse outputs [1], whereas GAN training is unstable and prone to mode collapse, reducing variety [1]. Diffusion is easier to condition on detailed text prompts, making it better for concept art from briefs, while GANs suit fast drafts or upscaling [1].

A compressed, organised representation of the images produced by the encoder [1], in which similar images are placed close together, so new images can be created by choosing or moving between points and decoding them [1].

Flow-based models learn an exact, invertible mapping between noise and images [1], so the process can be reversed without loss [1], allowing the studio to trace exactly how each image was formed and reproduce it precisely, which supports transparency and AI disclosure [1].

Strengths: combining models covers individual weaknesses. A VAE compresses images into a latent space so diffusion can run there, cutting computational cost (latent diffusion) and giving flexible editing for artistic exploration. Diffusion supplies photorealism, diversity and stable training. A GAN component sharpens detail and reduces generation time. A flow-based component adds reproducibility and traceability for transparency. Together they give higher quality, better control over outputs and more consistency for brand work.

Weaknesses: much more complex to design, train and maintain; needs specialist expertise the studio may not have; more components that can fail; combined training can still be unstable; higher development cost and possibly more infrastructure.

Judgement: a hybrid such as a latent diffusion model (VAE + diffusion) is worth it, as it directly addresses the studio's main challenge of computational demand while keeping quality. Adding GAN and flow components is justified only if the studio produces large volumes of branded content where speed and traceability matter; otherwise a pre-trained hybrid fine-tuned on its own licensed data is the best balance.

13. How to Get Full Marks

Command term	What to write
Define / State / Identify	One precise fact. Use the booklet's key words. No explanation needed.
Outline	A brief account: one point plus a little detail per mark.
Describe	A detailed account of what happens, in order. Count the marks, write that many steps.
Explain	Say why or how. Use "because", "so", "which means". Every point needs a reason.
Distinguish / Compare	Point by point, both sides in the same sentence. "Whereas" is your friend. Compare also needs at least one similarity.
Discuss / Evaluate / To what extent	Arguments for and against, each developed and applied to Visionary Studios, then a justified conclusion that actually answers the question.

Final checklist: you are ready when you can...

☐ Define every glossary term from memory.
☐ Draw and explain the forward and reverse diffusion process.
☐ Say what the CNN does and why diffusion is slow, with 4 reasons.
☐ Give 4 ways to reduce computational demand, each with a trade off.
☐ Compare text-to-image, conditional and unconditional, with a studio use for each.
☐ Explain character consistency and style adherence and how to achieve them.
☐ Explain dataset curation, bias and AI disclosure, each with a problem and a solution.
☐ Write an 8 mark "discuss" answer with a conclusion in under 15 minutes.
☐ HL Draw a GAN, explain the adversarial dynamic, mode collapse, instability and 3 mitigations.
☐ HL Explain VAEs, latent space and flow-based models in 2 sentences each.
☐ HL Design a hybrid pipeline for the studio and evaluate it.

Ninja Notes:

Quote the booklet's words: "iterative denoising", "noise injection", "coherent outputs", "adversarial dynamic", "mode collapse".
Always link to Visionary Studios: adverts, concept art, brand identity, recurring characters, clients, existing infrastructure.
Every advantage has a trade off. Examiners reward answers that show both.
Link to the syllabus: CNNs and training (A4.1), GPUs and cloud (A1.1, A4.1.3), ethics and bias (A4.4).
In long answers, finish with a clear judgement. "It depends" with no decision gets fewer marks.