Diffusion Models: Learning to Undo Noise · seed 1 · A4, ink-friendly. The answer key prints on its own page for grown-ups.
En anglais

Pictures from unmaking noise

Computing · Machine Learning · ages 22-24
Name ______________________   Date ____________

  1. What does the network predict at each step?

    • The noise mixed in so far
    • The clean image in a single jump
    • The number of steps still to run

  2. What does the forward process need training for?

    • Nothing, it is a fixed recipe
    • Learning how much noise to add at each step
    • Learning which pixels to corrupt first

  3. Generation starts from pure noise and denoises stepwise into a sample.

    Circle one:   True   False

  4. Why is stepwise denoising easier to learn than one-shot image making?

    • Each step is a small easy correction
    • Noise takes less memory than a full image
    • Long loops need no training data

  5. Samples show rough textures and odd speckles. Which fix needs no retraining?

    • Retrain with a different noise schedule
    • Add more layers to the network
    • Run more sampling steps

  6. Generation is too slow for a demo. What do you trade away by cutting steps?

    • Nothing, steps are decorative
    • The need for any model
    • Sample cleanliness

  7. Two configs give equal quality but one uses half the steps. Which ships?

    • The slower one, since more steps always mean cleaner samples
    • The half-step one, it buys the same quality faster
    • Neither, quality ties prove nothing

  8. A student trains the network to output clean images directly from pure noise. What is the conceptual error?

    • Networks cannot output images without captions
    • Direct jumps skip the easy stepwise structure
    • Noise must be added, never removed, during training
LightMySky · lightmysky.comW1-mt_kZdQJ96iZ3-s1

Answer key

For grown-ups. Fold this page away before handing over the rest.

Pictures from unmaking noise W1-mt_kZdQJ96iZ3-s1

  1. The noise mixed in so far · Predicting the mixed-in noise lets generation remove part of it at each step.
  2. Nothing, it is a fixed recipe · Corruption just adds noise by schedule.
  3. True · Running the chain backwards builds the sample.
  4. Each step is a small easy correction · Small corrections beat one giant leap.
  5. Run more sampling steps · More steps are a sampling-time setting that needs no retraining. They usually help when the cause was too few steps, while the other two fixes need a new training run.
  6. Sample cleanliness · Fewer steps run fast but usually leave rougher samples.
  7. The half-step one, it buys the same quality faster · Steps are a cost, so equal quality at fewer steps wins.
  8. Direct jumps skip the easy stepwise structure · Diffusion works because each step is a small learned undo.
Worksheet · LightMySky