IB CS 2027 case study practice questions

Question 1

Hearth & Oak is deciding whether to launch its "Design my room" feature using a diffusion model trained only on licensed images and its own product photos. To what extent would this be a good decision? In your answer, consider image quality, speed and ethical issues.

Accepted Answer

A diffusion model is a sensible choice for quality: it produces realistic, varied rooms and follows a customer's description well, and fine-tuning on Hearth & Oak's own studio photos helps it show the actual products accurately. A dataset of only licensed images is smaller, though, so the model may handle unusual requests ("a houseboat with a velvet sofa") poorly.

Speed is the main weakness. Each image needs many denoising passes, so a customer may wait, and every press of the button costs GPU time. Running diffusion in latent space and using a distilled model with only a few steps would bring this down to a few seconds.

Ethically, licensed data is a strong choice: Hearth & Oak can show it had permission for everything it trained on, which protects a brand built on trust. It still needs to check its rooms are not all one style of home and to label images as AI-generated so customers do not mistake a scene for a photo.

Overall it is a good decision to a large extent, provided Hearth & Oak uses a fast, distilled model so the feature responds quickly, and labels the images clearly.

Question 2

Hearth & Oak has three options for generating catalogue scenes and live "Design my room" images: (A) a GAN, (B) a diffusion model, (C) a hybrid of latent diffusion with a distilled few-step model for live use. Evaluate the three options and recommend one.

Accepted Answer

A GAN would be the fastest and cheapest option: it makes an image in a single pass, so live previews would feel instant and cost little. But GANs often collapse to a narrow range of outputs, so different customers would see suspiciously similar rooms, and their adversarial training is unstable and needs expert tuning that a small team may not have. They also follow free-text prompts less well than diffusion models.

A plain diffusion model gives the best quality and variety, follows descriptions closely through cross-attention, and trains stably with a simple denoising loss. For the catalogue, where 3,000 scenes are made once in overnight batches, its slowness does not matter. For live use it does: dozens of full denoising passes per image would make customers wait and multiply GPU costs every time anyone presses the button.

The hybrid uses each part where it is strongest. Working in latent space cuts the cost of every step, the full model can produce catalogue scenes at top quality, and a distilled few-step model can serve live requests in a second or two. Its weaknesses are complexity, with more parts to train, test and update, a distilled model that may be less varied than its teacher, and an autoencoder decoder that can soften fine product details such as fabric.

Weighing these, the GAN's speed advantage is outweighed by its variety and training problems, which hit a small team hardest, and plain diffusion is too slow for the live feature. Hearth & Oak should choose the hybrid, but adopt an existing, well-supported latent diffusion model and its distilled version rather than building one, so the complexity is mostly someone else's, and composite the real product photo into each scene so that what customers see is exactly what they will buy.

Question 3

Hearth & Oak wants customers to get the same room image if they press "Generate" twice with the same description. Explain how a diffusion model can produce the same image twice, and why it usually produces a different one.

Accepted Answer

A diffusion model starts every generation from an image of random noise and removes noise from it step by step. If the starting noise is different, the model is guided to a different final image, so pressing "Generate" twice normally gives two different rooms. The starting noise is produced from a seed. If Hearth & Oak stores the seed and uses it again with the same prompt and the same sampler settings, the starting noise is identical, and because the seed also fixes any randomness used during the denoising steps, the final image is the same.

Question 4

Outline the forward process in a diffusion model.

Accepted Answer

The forward process takes a training image and adds a small amount of Gaussian noise to it over many steps, following a fixed noise schedule, until the image is indistinguishable from pure random noise. Nothing is learned in this process.

Question 5

Describe how the denoising network in a diffusion model is trained.

Accepted Answer

For each training example, a real image is taken and a random step number is chosen. Noise is added to the image to match that step, and the exact noise is remembered. The network is given the noisy image and the step number and predicts the noise. The prediction is compared with the real noise using a loss such as mean squared error, and backpropagation with gradient descent adjusts the weights to reduce the error. This is repeated over millions of images and noise levels.

Question 6

Customers using Hearth & Oak's "Design my room" box complain that they wait too long for an image. Suggest two changes to the diffusion system that would reduce the waiting time, and state a disadvantage of each.

Accepted Answer

First, Hearth & Oak could reduce the number of sampling steps, for example from 50 to 20, using a sampler designed for few steps. Each step is a full pass through the network, so this cuts the time roughly in proportion, but images may lose fine detail such as fabric texture. Second, it could generate at a lower resolution and enlarge the result. Fewer pixels means less computation per step, but the enlarged image may look soft, which matters when customers are judging a product's material.

Question 7

Explain why a U-Net denoiser uses skip connections.

Accepted Answer

The encoder half of a U-Net repeatedly downsamples the image, which captures what is in the picture but loses exactly where edges and small details are. Skip connections copy the feature maps from each encoder level straight across to the decoder level of the same size. The decoder can then combine the overall understanding from the bottleneck with the precise detail from early layers, which matters because the denoiser must predict noise for every single pixel.

Question 8

Define the term "kernel" in the context of a convolutional neural network.

Accepted Answer

A kernel is a small matrix of learned weights that is moved across an image, multiplying and summing the values under it, to detect one kind of feature and produce a feature map.

Question 9

Explain why a convolutional layer is more suitable than a fully connected layer for processing images.

Accepted Answer

An image has hundreds of thousands of pixel values, so a fully connected layer would need a weight between every pixel and every neuron, which is too many to train or run efficiently. A convolutional layer instead slides one small kernel over the image and reuses the same weights everywhere, so it has far fewer parameters. Each kernel only looks at a small neighbourhood, which suits images because nearby pixels form edges and textures. Because the same kernel is used everywhere, a pattern it learns, such as the edge of a table leg, is recognised wherever it appears.

Question 10

Hearth & Oak's generated rooms often have furniture with warped legs and mismatched edges. Suggest how the structure of a U-Net relates to this problem.

Accepted Answer

Straight table legs and clean edges are fine detail that has to be correct at full resolution. In a U-Net the middle of the network sees a heavily downsampled version of the image, so it knows roughly where a table is but not exactly where its edges are; that precision has to come back through the skip connections and the high-resolution layers. If those layers are small, or the model has seen few clean product photos, edges may come out warped. Hearth & Oak could fine-tune on its own sharp studio photos or generate at a higher resolution.

Question 11

Hearth & Oak is choosing between a model trained on images scraped from the web and a model trained only on licensed stock photos and its own studio photos. Discuss which it should choose.

Accepted Answer

The web-scraped model has learned from far more images, so it is likely to produce more realistic and varied rooms, from tiny city flats to farmhouse kitchens. However, the photographers and designers whose work it learned from did not agree and are not credited or paid. That is an ethical problem, and also a business one: Hearth & Oak could face copyright claims, especially if the model reproduces a near-copy of a well-known interior photo, and a story about "stolen" images would damage a brand that depends on trust.

The licensed model avoids most of this. Hearth & Oak can show where every training image came from, and the generated images are safer to use in adverts. The cost is a smaller, narrower dataset, so rooms may look less varied, and licences cost money.

On balance Hearth & Oak should choose the licensed model. Its main need is realistic scenes around its own products, which its studio photos already cover well, and the legal and reputational risks of the scraped model are much harder to fix later than a narrower range of room styles.

Question 12

Outline what is meant by memorisation in a generative image model.

Accepted Answer

Memorisation is when a generative model outputs a copy or near-copy of one of its training images instead of a new image. It is more likely for images duplicated many times in the training set, and it can expose copyrighted or private pictures.

Question 13

Identify two ways Hearth & Oak could make it clear to customers that a room image was generated by AI.

Accepted Answer

It could show a visible "AI-generated scene" label on every generated image, and attach provenance metadata (content credentials) to the file that records that it was made by an AI model.

Question 14

A customer types "a living room in the style of [a named living illustrator]" into Hearth & Oak's "Design my room" box. Explain one ethical issue this raises and one technical measure Hearth & Oak could take.

Accepted Answer

The prompt asks the model to imitate a specific living illustrator. That artist did not agree to have their work used this way and gets nothing from it, and the result could compete with the commissions they rely on or be passed off as theirs. Hearth & Oak could check prompts against a list of living artists' names and refuse or rewrite those requests. This stops the most direct cases, although a determined user could still describe the style in other words, so it reduces the problem rather than removing it.

Question 15

Almost every room Hearth & Oak's model generates is large, bright and minimalist, even when the customer asks for "a cosy family home". Explain how the training data could have caused this, and suggest one way to reduce it.

Accepted Answer

Most interior photos online are taken by professionals to sell homes or furniture, so they show large, bright, tidy rooms. If the model was trained mostly on such photos, it learned that this is what "a room" looks like and makes that by default, even pulling a "cosy family home" towards it. Hearth & Oak could add and upweight photos of smaller, lived-in homes, including its own customers' photos shared with permission, and then generate a few hundred images for prompts like "a cosy family home" to check that the variety has improved.

Question 16

State two sources of bias in the training data of an image generation model.

Accepted Answer

Some groups and places appear far more often than others in images collected from the web, and the captions attached to images carry the assumptions of the people who wrote them.

Question 17

Explain why removing duplicate images from a training dataset is good practice.

Accepted Answer

Images that appear many times are much more likely to be memorised and copied in the output, and duplicates make popular images count more than others, skewing what the model learns as normal.

Question 18

Hearth & Oak plans to filter its training data by removing every image an automatic quality model scores as "low quality". Discuss this plan.

Accepted Answer

Filtering out low-quality images should make the model's rooms sharper and better lit, which matters for a shop whose images must show products accurately, and it saves training time on blurry or irrelevant pictures. However, the quality model was itself trained on someone's idea of a "good" photo. It may score phone photos of small, cluttered or older homes as low quality far more often than professional shots of large ones. Removing them would make Hearth & Oak's generator even worse at showing the homes many customers actually live in. Hearth & Oak should still filter, but first check a sample of what the filter removes, broken down by kind of room, and lower the threshold or keep a set of real-home photos if one kind is being removed much more than others.

Question 19

Describe how the generator in a GAN is trained.

Accepted Answer

The generator is given random latent vectors and produces a batch of images. These are fed into the discriminator, which outputs how likely each is to be real. The generator's loss is low when the discriminator is fooled, so the error is backpropagated back through the discriminator into the generator, and only the generator's weights are updated while the discriminator's are held fixed. This step alternates with a step that trains the discriminator on real and generated images.

Question 20

State the role of the discriminator in a GAN.

Accepted Answer

The discriminator classifies an image as real, from the training data, or fake, made by the generator.

Question 21

Hearth & Oak wants to show live previews that update as a customer drags a slider for "how modern" a room is. Explain why a GAN might be more suitable than a diffusion model for this feature.

Accepted Answer

A GAN makes an image with one pass through its generator, while a diffusion model needs many passes, one per denoising step. A live preview has to produce a new image every time the slider moves, many times a second, which a single pass can manage and many passes cannot. A GAN's latent space also tends to change smoothly, so small slider moves give small, gradual changes in the room.

Question 22

Compare and contrast GANs and diffusion models as approaches to image generation.

Accepted Answer

Both are neural networks that turn random noise into new images after training on large datasets. A GAN produces an image in one pass of its generator, while a diffusion model removes noise over many steps, so GANs are much faster at generation time. GANs are trained adversarially against a discriminator, which is unstable, whereas diffusion models are trained with a simple, stable denoising loss. Diffusion models usually produce more varied images and follow text prompts better, while GANs can collapse to a small range of outputs.

Question 23

Hearth & Oak's GAN produces beautiful sofas, but testers notice that almost all of them are the same shape and colour. Identify this problem and explain one technique that could reduce it.

Accepted Answer

This is mode collapse. The generator has found a few sofa images that the discriminator accepts as real, and because its only goal is to fool the discriminator, nothing rewards it for making other shapes or colours. Minibatch discrimination lets the discriminator compare the images within a batch rather than judging each alone. A batch of near-identical sofas then looks suspicious, because real batches are varied, so the generator is pushed to spread its output across more kinds of sofa.

Question 24

Explain why a discriminator that is too strong can stop a GAN from learning.

Accepted Answer

If the discriminator becomes too good, it labels every generated image as fake with near certainty. Small changes to the generator then make almost no difference to the loss, so the gradient passed back to the generator is close to zero. With almost no signal to learn from, the generator's weights barely change and it stops getting better.

Question 25

Outline two reasons why the loss values of a GAN are a poor guide to whether training is going well.

Accepted Answer

First, each network is judged against the other, and the other keeps changing, so a lower generator loss might only mean the discriminator has become weaker rather than the images better. Second, the losses measure how well each network does in the game, not how varied the images are, so the generator can collapse to a few outputs while the losses look fine. Teams therefore look at sample images and a score such as FID.

Question 26

Hearth & Oak's engineers have limited time. Evaluate replacing the standard GAN loss with a Wasserstein loss.

Accepted Answer

A Wasserstein loss measures how far apart the real and generated distributions are with a smooth score, so the generator still gets a useful gradient even when the critic is strong. Its value also tracks image quality more closely, which would help Hearth & Oak's small team see whether training is improving. However, it needs a gradient penalty, which adds computation to each step, and typically several critic updates per generator update, so training is slower and still needs some tuning. For a team with little time to babysit training runs, it is worth it: fewer failed runs saves more time than the slower steps cost.

Question 27

Explain why most text-to-image diffusion systems run the diffusion process in a latent space rather than on the pixels of the image.

Accepted Answer

A variational autoencoder first compresses the image into a latent grid many times smaller than the pixel image, and a decoder turns latents back into pixels. Running diffusion on the latent means each denoising step processes far fewer values, so it needs much less computation and memory. Because generation repeats that step many times, the saving is multiplied, making both training and generation much faster and cheaper. The latent keeps the important structure of the image, so quality stays high, although very fine detail now depends on how good the decoder is.

Question 28

Outline one strength and one weakness of a variational autoencoder (VAE) as an image generator.

Accepted Answer

A VAE generates an image in one fast pass of its decoder and trains stably, but its images tend to be blurrier and less detailed than those from GANs or diffusion models.

Question 29

Hearth & Oak is considering adding a GAN-based upscaler after its diffusion model, so it can generate at a lower resolution and enlarge the result. Discuss this design.

Accepted Answer

Generating at a lower resolution would cut the time and GPU cost of each diffusion run a lot, and a GAN upscaler adds sharp texture in one quick pass, so customers would get crisp images faster. The danger for Hearth & Oak is that an upscaler invents plausible detail. A fabric weave or wood grain that the upscaler made up could make a sofa look unlike the one the customer will receive, which breaks the trust the brand depends on. It is also one more model to train, test and keep updated. A good compromise is to upscale only the room around the product and composite the real studio photo of the product on top. With that safeguard the design is worth adopting; without it, the risk of misleading customers outweighs the speed.

Question 30

Distinguish between a diffusion model and a flow-based model.

Accepted Answer

A diffusion model is trained to reverse a noising process approximately, by removing a little predicted noise at each of many steps, whereas a classic flow-based model is built from exactly invertible transformations, so it can turn noise into images and images back into noise and calculate exactly how likely an image is.

Question 31

Hearth & Oak needs two kinds of images: room scenes for 3,000 catalogue pages, made once, and live images for the "Design my room" box. Suggest suitable settings for each, justifying your choices.

Accepted Answer

For the catalogue, Hearth & Oak should use the highest quality settings: a large model, many sampling steps and a high resolution. These images are made once, can be generated overnight in large batches when GPUs are cheaper, and appear on every product page, so quality matters more than time. Staff can review them and regenerate any with mistakes before they go live.

For the "Design my room" box, the customer is waiting, so it should use a distilled model or far fewer steps at a lower resolution, aiming for an image in a few seconds. This also matters for cost, because it runs every time any customer presses the button. It could show a quick preview first and refine only the image the customer likes, and label the result as an impression of the room.

Question 32

Distinguish between latency and throughput for an image generation service.

Accepted Answer

Latency is how long one customer waits from sending a prompt to receiving the image, while throughput is how many images the whole service can produce per second or per hour.

Question 33

Explain one limitation of using FID alone to compare two image generation models.

Accepted Answer

FID only compares the overall distribution of generated images with real ones, so it does not measure whether an image matches its prompt. A model could get a good FID while ignoring what customers asked for, so it should be combined with a prompt-matching measure such as CLIP score and human review.

Question 34

Hearth & Oak could run its image model on its own rented cloud GPUs, or pay a third-party image generation service per image. Evaluate these two options.

Accepted Answer

Running its own model gives Hearth & Oak full control: it can choose a model trained on licensed data, fine-tune it on its own product photos, and keep customers' prompts in-house. But it needs engineers who can run GPU servers, it pays for GPUs even when nobody is generating, and it must cope with busy sale weekends itself. A third-party service needs no setup, scales automatically and charges only for images made, which suits a small team launching a new feature. However, Hearth & Oak cannot see or choose what that model was trained on, so the ethical and copyright questions are out of its hands, and customer prompts leave the company. Per-image pricing is cheaper while volumes are low, but at large, steady volumes renting its own GPUs usually costs less. Hearth & Oak should start with a service that can show its training data is licensed, measure demand, and move to its own GPUs only if volume grows enough to justify the engineering work.

Case study practice questions

The company in every question

Extended questions

How diffusion models make an image

Training data: ethics, ownership and bias

GANs: a generator against a discriminator

Hybrid models and the trade-offs between them