DDPM and DDIM Samplers Quiz

Answers and explanations appear as you take the quiz.

1. What is the main practical advantage of DDIM over DDPM?
   • It can produce comparable images in far fewer sampling steps
   • It requires a completely different trained network
   • It adds more random noise at each step for better quality
   • It only works on black-and-white images
2. What does a diffusion model's neural network actually learn to predict during training?
   • The final caption for the image
   • The noise that was added to an image at a given timestep
   • The compression ratio of the image
   • The camera settings used to take the photo
3. Why can DDIM produce the exact same image from the same starting noise every time?
   • Because it caches previously generated images
   • Because it always uses a fixed prompt
   • Because its sampling process is deterministic with no injected randomness
   • Because it averages many random runs together
4. In DDIM, what value of the parameter eta recovers the original stochastic DDPM behavior?
   • eta = 0
   • eta = -1
   • eta = 100
   • eta = 1
5. Roughly how many steps did the original DDPM typically need for high-quality samples?
   • Hundreds to about a thousand steps
   • Exactly 3 steps
   • A single step
   • Over one million steps
6. Which of these is a key intermediate quantity DDIM estimates before stepping forward?
   • The file size of the output
   • The predicted clean image x0
   • The user's prompt embedding length
   • The GPU temperature
7. What does DDIM change compared with DDPM when using a model trained with the DDPM procedure?
   • It replaces the network with a text classifier
   • It changes the training data from images to audio
   • It changes the sampling trajectory while reusing the trained denoising network
   • It removes the need to estimate noise
8. Why can DDIM skip selected diffusion timesteps during sampling?
   • The training process never added noise
   • Each skipped timestep is replaced by a larger image
   • It copies an earlier generated image from a cache
   • Its non-Markovian sampling path can connect noise levels without visiting every training step