Skip to content
ExamSpine

Trade-offs: quality, speed, cost and complexity

Every setting in an image generator buys something and costs something. The exam rewards weighing them for a specific company, not listing them.

Challenge 4HL only

The settings that move the trade-off

SettingRaise it and you getBut you pay
Sampling stepsMore detail, fewer artefacts (up to a point)Time and GPU cost, roughly in proportion
ResolutionSharper, larger imagesMuch more computation (the work grows faster than the width)
Model sizeBetter quality and prompt followingMore GPU memory, slower, costlier
Guidance scaleCloser match to the promptLess variety, oversaturated images if too high
Batch sizeMore images per second per GPU (throughput)Each customer may wait longer (latency)

Measuring quality

  • FID (Fréchet Inception Distance). Features of many real and many generated images are extracted with a pretrained image network and the two distributions compared. Lower is better. It reflects both quality and variety, but not whether images match their prompts.
  • CLIP score. Measures how well an image matches its text prompt using a model trained on image and caption pairs. Higher is better.
  • Human evaluation. People compare images side by side. Slow and costly, but it measures what customers actually care about.
  • Task-specific checks. For a shop: is the product shown correctly? Are its proportions right? A general score will not tell you.

Key idea: Automatic scores are cheap and repeatable, but only humans can tell you whether an image is good for this purpose.

Speed and cost in deployment

  • Latency is how long one user waits; throughput is how many images a system produces per second. Batching improves throughput but can worsen latency.
  • Generation usually runs on GPUs in the cloud, paid per second of use. Cost per image is roughly time per image times price per GPU-second, so halving the steps roughly halves the cost.
  • Ways to cut cost: fewer steps with a better sampler, distilled models, lower precision numbers (quantisation), caching popular results, generating overnight for catalogue images rather than live.
  • Running on the user's own device saves server cost and keeps prompts private, but most phones are too slow for large models.

A frame for any evaluation question

  1. What does this company need most? (A shop: accurate products and trust. A game studio: variety and style.)
  2. Option A: two strengths, two weaknesses, each linked to that need.
  3. Option B: the same.
  4. Weigh: which weaknesses matter most here, and can any be reduced?
  5. Conclude: a clear choice, with the deciding reason, and a condition ("unless the budget allows...").

Key idea: The conclusion earns marks only if it follows from the points you made.

Key terms

Latency
How long one request takes, from sending a prompt to getting the image.
Throughput
How many images a system produces in a given time. Batching raises it, sometimes at the cost of latency.
FID (Fréchet Inception Distance)
A score comparing features of many real and many generated images. Lower means the generated set is closer to the real one.
CLIP score
A measure of how well an image matches a text prompt, using a model trained on image and caption pairs. Higher is better.
Guidance scale
How strongly classifier-free guidance pushes an image towards the prompt. Higher follows the prompt more but gives less variety.
Quantisation
Storing a model's weights as lower-precision numbers so it runs faster and uses less memory, with a small loss of quality.
GPU
A graphics processing unit: a processor with thousands of simple cores that does the parallel arithmetic of neural networks quickly.
Sampling steps
The number of denoising passes used to generate one image. More steps is slower but can add detail.

See all 65 terms in the glossary

Exam-style questions

In the style of Paper 1 Section B. Write your answer first, then open the marking points and mark yourself honestly. The questions use our own company, Hearth & Oak: an online furniture shop that wants to generate room scenes around its products.

Worked example · Suggest · AO2 · HL

[6 marks]

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.

How to approach it

"Suggest" with "justifying": every setting needs a reason tied to the situation. Two situations, so split the marks evenly and contrast them: made once vs made on demand, quality-critical vs latency-critical.

0 words

Marking points and model answer

A model 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.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: Catalogue: many sampling steps / high resolution / the largest model. (1 mark)
  • Point 2: Justified: made once and not time-critical (can run overnight / in batches), and quality matters for every page. (1 mark)
  • Point 3: Catalogue: humans can review and regenerate poor images before publishing. (1 mark)
  • Point 4: Live: few steps / distilled or smaller model / lower resolution. (1 mark)
  • Point 5: Justified: customers wait for it, so latency matters; cost is paid every time a customer generates. (1 mark)
  • Point 6: Live: a safeguard, for example a preview then refine, caching, or labelling the image as an impression. (1 mark)

Earns nothing

  • The same settings for both with no justification.

Mark every point to see your total.

Question 1 · Distinguish · AO1 · HL

[2 marks]

Distinguish between latency and throughput for an image generation service.

0 words

Marking points and model answer

A model 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.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: Latency: the time one request takes, from prompt to image. (1 mark)
  • Point 2: Throughput: how many images the system produces in a given time. (1 mark)

Mark every point to see your total.

Question 2 · Explain · AO1 · HL

[2 marks]

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

0 words

Marking points and model answer

A model 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.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: A limitation, for example it does not measure how well images match their prompts / it depends on the pretrained network's features / it needs many images / it does not check task-specific correctness. (1 mark)
  • Point 2: Why it matters, for example a model could score well while ignoring prompts or showing products wrongly, so human or CLIP checks are also needed. (1 mark)

Mark every point to see your total.

Question 3 · Evaluate · AO3 · HL

[6 marks]

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.

0 words

Marking points and model answer

A model 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.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: Own GPUs: full control over the model (can fine-tune on its products, choose training data), and its data and prompts stay with it. (1 mark)
  • Point 2: Own GPUs: needs expertise and maintenance; pays for GPUs even when idle; scaling for busy periods is its job. (1 mark)
  • Point 3: Service: no setup, scales automatically, pay only per image; quick to launch for a small team. (1 mark)
  • Point 4: Service: less control (cannot choose training data, so ethical/IP questions are out of its hands), customer prompts sent to a third party, cost per image can rise, dependency on supplier. (1 mark)
  • Point 5: Links cost to volume: per-image pricing is cheaper at low volume; own GPUs can be cheaper at high, steady volume. (1 mark)
  • Point 6: A justified conclusion for this company. (1 mark)

Mark every point to see your total.

Glossary

65 key terms, in plain English

Practice questions

Paper 1 Section B style, with full answers

All case study topics

  1. Challenge 1

    SL and HL

    How diffusion models make an image

    Noise, denoising, the network that predicts the noise, and how a text prompt steers it.

  2. Challenge 2

    SL and HL

    Training data: ethics, ownership and bias

    Consent, copyright and attribution for training images, and how curation choices shape what the model makes.

  3. Challenge 3

    HL only

    GANs: a generator against a discriminator

    How two networks train against each other, and why that training is unstable.

  4. Challenge 4

    HL only

    Hybrid models and the trade-offs between them

    Systems that combine diffusion, GANs and autoencoders, judged on quality, speed, diversity and complexity.

Written by Spine in our own words. Spine is not affiliated with or endorsed by the International Baccalaureate. Every question, set of marking points and model answer here is our own, set in our own made-up company. Facts about the case study itself come from public guides and are marked “reported”. Last checked 2026-10-23.