The scenario

Visionary Studios, a (fictional) creative design company, is evaluating generative AI models to produce high-quality images faster for advertising campaigns, concept art and digital media. It is comparing diffusion models, GANs and hybrid models, and has to weigh output quality, cost, reliability and the ethical and legal risks.

Read the booklet itself first: Case study booklet (PDF). These notes explain every idea in it, go a bit beyond it (the exam expects your own research), and link to the activities.

What SL and HL need

Sections in these notes marked HL are for HL only (use Hide HL in the menu to hide them). The exam can also give you new information about the scenario and ask you to apply what you know to it, so understand the ideas rather than memorising the booklet.

How the pieces fit together

Text-to-image generation

• The model creates an image from a text prompt, e.g. "a fox in a red coat, watercolour". Examples: DALL·E, Stable Diffusion, Midjourney.
• A text encoder turns the prompt into an embedding (a vector of numbers that represents its meaning). The embedding conditions (guides) every denoising step, so the image matches the words.
• Useful for concept art and advertising, where a specific theme, product or style is required, and for producing many variations quickly.
• Text-to-image is a kind of conditional generation: the condition is the text.

Conditional image generation

The model generates an image from a specific input (the condition), so the output matches a predefined objective or style.

Unconditional image generation

• No input except random noise: the model produces images that look like its training data, relying only on the patterns it learned.
• Useful for exploratory creativity, abstract art and unexpected visuals.
• Valuable for synthetic datasets: generating extra training images for other AI models (e.g. rare cases). It reduces the direct use of real people's photos, although the generator itself was trained on real images and can memorise them.
• Downside: no control over what appears, so it is less useful for a client brief.

The big idea

A diffusion model learns to remove noise. If a network can turn a slightly noisy image into a slightly cleaner one, then by repeating that many times it can turn pure random noise into a brand-new image. This is why diffusion is called iterative.

Training: use the forward (noising) process to make examples, then learn to predict the noise

1. Take a real image from the training dataset.
2. Pick a random step t (e.g. between 1 and 1000) and add the matching amount of Gaussian noise (noise injection). The computer knows exactly what noise it added.
3. Show the noisy image (and t, and the prompt for a conditional model) to the neural denoiser, usually a CNN called a U-Net, and ask it to predict the noise.
4. Compare its prediction with the real noise (the loss) and adjust the weights by backpropagation.
5. Repeat millions of times with different images and amounts of noise.

Generation (the reverse process): denoise step by step

1. Start from a random noise image (a new random seed gives a different picture).
2. The CNN predicts the noise in the current image (guided by the prompt embedding if there is one).
3. Subtract part of the predicted noise to get a slightly cleaner image.
4. Repeat for every step (e.g. 1000 in the original method, typically 20–50 with modern samplers) until a clear image remains.

The dataset decides the content: a denoiser trained on animal photos "finds" cats and dogs in the noise; one trained on cities produces streets and skylines. It can only create what resembles what it has learned.

DDPM and the CNN denoiser

• DDPM (denoising diffusion probabilistic model, Ho et al., 2020) is the method that formalised this: a fixed schedule for gradually adding noise in training and a learned model for gradually removing it in generation. Modern diffusion models are based on it.
• It is a probabilistic model: generation involves randomness (the random starting noise and, in DDPM, a little fresh noise at each step), so a new random seed gives a different image for the same prompt; the same seed and settings reproduce the same image.
• The denoiser is a convolutional neural network: small filters slide over the image to detect features (edges, textures, then shapes and objects), which is ideal for images because the same feature can appear anywhere. (CNNs are an HL syllabus topic, A4.3.9; SL students only need this basic idea for the case study.) The booklet says the denoiser is "typically" a CNN: the U-Net used by Stable Diffusion 1 and 2 and many others. Some newer models, such as Stable Diffusion 3, use a transformer instead.

Why diffusion is expensive

• Many steps: each image needs tens to hundreds of passes through the denoiser.
• A big network: the CNN has hundreds of millions to billions of parameters, so each pass is a huge number of multiplications.
• High resolution: a 512×512 colour image has 786,432 values; a 1024×1024 image four times as many. Photorealism needs high resolution.
• Training on hundreds of millions of image–caption pairs: the original Stable Diffusion v1 used about 150,000 GPU-hours on high-end data-centre GPUs.
• Roughly: cost ∝ steps × cost of one pass × number of images. A studio generating thousands of variations a day pays for every step.

Ways to manage it

Optimising the process is a balance between speed, cost and image quality. A good exam answer names the technique, explains why it saves computation and states the trade-off.

Character consistency: embedding-based approaches

• Problem: generate the same mascot in 20 different ad scenes. Each generation starts from new random noise, so by default the character's face, clothes and colours drift.
• An embedding is a vector of numbers that represents something (a word, an image, a character). If the character is captured as a fixed embedding and the same embedding is fed in every time, the model reproduces the same character while the rest of the prompt changes the scene.
• Ways to do it: learn a new "word" for the character from a few reference images (textual inversion), fine-tune the model on the character (e.g. DreamBooth, LoRA), or feed in a reference image's embedding at generation time (image-prompt adapters). Reusing the same random seed helps only for small changes to the same prompt.

Style adherence

• Conditional models (like text-to-image) must produce the style or theme the prompt asks for, e.g. the brand's flat pastel illustration style.
• Techniques: a guidance scale that controls how strictly the image follows the prompt (higher = closer to the prompt but less variety); negative prompts ("no photo, no 3D"); fine-tuning on the brand's own artwork; structural conditions such as sketches or segmentation maps.

Dataset curation and intellectual property

• Many models were trained on billions of images scraped from the web (e.g. the LAION datasets), including copyrighted photos and artwork, usually without the creators' permission.
• Risks for the studio: legal claims, outputs that closely copy a training image or reproduce a watermark, logo or famous character, and damage to reputation with the artists it works with.
• Real case: Getty Images v Stability AI (UK High Court, November 2025). Getty dropped its main claim about training, and its secondary copyright claim failed; it won only a very limited trade mark claim about Getty watermarks appearing in some outputs. Getty was later given permission to appeal, so the law is still developing, and similar cases are under way in other countries.
• Curation means deliberately choosing and checking the training data: use licensed, public-domain or the studio's own images (Adobe says its Firefly models were trained on licensed and public-domain content); respect artists' opt-outs; remove duplicates, watermarks and personal images; keep records of where every image came from.
• Trade-off: a smaller, carefully curated dataset is safer but more expensive to build and may give less variety.

Bias and fairness

• A model learns the patterns in its data, including its imbalances, and can exaggerate them. A Bloomberg analysis (2023) of over 5,000 Stable Diffusion images found that images for high-paying jobs were dominated by lighter-skinned men, and low-paying jobs by darker-skinned people and women.
• For a studio this means exclusionary or inaccurate adverts, offended audiences and clients, and possible discrimination complaints.
• Bias mitigation: audit the dataset (who and what is represented), rebalance or add data, test outputs across many prompts and count who appears, human review before publishing, write prompts that specify diversity where appropriate, and document the model's limitations.

Transparency and AI disclosure

• Clients and audiences should know when an image is AI-generated: this maintains trust and avoids misleading people (e.g. a "photo" of a product that does not exist in that form).
• Methods: visible labels, contract terms with clients, invisible watermarks, and content credentials (the C2PA standard attaches signed information about how an image was made).
• The law is catching up: the EU AI Act (Article 50) has applied since 2 August 2026. Providers of generative AI must mark outputs in a machine-readable way (systems already on the market before then have until 2 December 2026), and people who publish deepfakes must label them clearly.

Other issues worth researching