AIF-C01 Practice Question: Is a key advantage of using a diffusion model for…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Diffusion models produce more diverse and higher-quality images with stable training

  Why this is correct

  Diffusion models iteratively denoise random noise, avoiding the adversarial min-max game that destabilises GAN training. This yields greater sample diversity and higher fidelity, directly satisfying the stem's demand for a key advantage over GANs. Mode collapse, a common GAN failure, is largely absent, giving the stable training the option claims.

• ✗

  Diffusion models generate images faster than GANs during inference

  Why it's wrong here

  Diffusion sampling iterates over many denoising steps, so inference is typically slower than a GAN's single forward pass through the generator. It is tempting because diffusion training is stable and parallelisable, but the question asks about inference speed, where GANs retain the advantage.

• ✗

  Diffusion models are inherently conditional and do not require labels

  Why it's wrong here

  Diffusion models are not inherently conditional; class or text conditioning is added through guidance mechanisms, and training still uses labelled or paired data. It is tempting because classifier-free guidance feels intrinsic, but conditioning is an added design choice, not a property that removes the need for labels.

• ✗

  Diffusion models require less training data than GANs

  Why it's wrong here

  Diffusion models generally need large datasets to learn the denoising distribution well, often more than a GAN. It is tempting because their stable training avoids mode collapse, which can look like data efficiency, but the advantage lies in training stability and sample diversity, not reduced data requirements.

Go deeper

Related to this question