Diffusion Models

Answer explanation

• They are unconditionally independent.

• They are always dependent.

• None of the above

Answer explanation

Why do latent diffusion models often perform better than standard VAEs for image

generation?

• VAEs use simple Gaussian priors while diffusion models use learned priors.

• VAEs have a sampling distribution mismatch between training and generation.

• Diffusion models can handle higher-dimensional data more efficiently.

• AEs cannot capture complex image distributions.

Answer explanation

In the continuous-time limit of diffusion models, what is the correct form of the

backward SDE?

• 
• 
• 
• 

Answer explanation

• Option 1

• Option 2

• Option 3

Answer explanation

This is just a silly question so don't worry if you got it wrong. A keen eye however would have noticed that only option 2 was Gaussian noise! It is worth noting at this point that different degradation methods can also work: https://arxiv.org/pdf/2208.09392, but we are primarily interested in the vanilla diffusion models.

• 
• 
• 
• 

Answer explanation

What makes training latent diffusion models computationally more efficient than

training standard diffusion models?

• They use fewer denoising steps.

• They have simpler network architectures.

• They require less training data.

• They operate in a lower-dimensional latent space.

Answer explanation

• 
• 
• 
• None of the above

Answer explanation