VAE and Diffusion Models

Question 1 of 5

Why does a VAE need a KL divergence term in its loss function, even though it already includes a reconstruction error?

To prevent the encoder from collapsing to a deterministic autoencoder.To allow the decoder to produce multiple reconstructions for the same input.To reduce the computational cost of sampling latent codes.To make the loss function differentiable with respect to the latent variables.

Question 2 of 5

What problem does the reparametrization trick solve in training a VAE?

It allows gradients to flow through the random sampling step.It ensures the encoder outputs a valid Gaussian distribution by making the variance positive.It reduces the KL divergence between the approximate posterior and the prior by enabling a closed-form expression.It makes the training objective convex, so gradient descent always converges to the global minimum.

Question 3 of 5

Why is the VAE loss derived from the ELBO considered principled rather than ad‑hoc?

It naturally arises from the probabilistic model as the only way to combine reconstruction and regularization.It uses a closed-form KL divergence, which is computationally cheaper than sampling-based methods.It minimizes the reconstruction error directly, so it is guaranteed to produce sharp images.It doesn't require the encoder to predict a variance, so it has fewer parameters than a standard autoencoder.

Question 4 of 5

How does DDIM achieve faster generation compared to DDPM?

It defines a non-Markovian forward process with the same marginals, enabling larger deterministic steps during reverse.It uses a pretrained autoencoder to compress images before applying diffusion, so each step processes lower-dimensional data.It replaces the Gaussian diffusion process with a deterministic ODE, which can be solved in fewer steps.It trains the denoising network to directly predict x_0 instead of the noise, making each reverse step simpler.

Question 5 of 5

Why is Stable Diffusion more computationally efficient than pixel-space diffusion models?

It performs diffusion in a lower-dimensional latent space obtained from a pretrained autoencoder, drastically reducing data size.It uses a transformer-based denoising network, which is more parallelizable than U-Nets.It avoids the need for a noise schedule by learning the reverse process implicitly, skipping many steps.It uses an ODE solver with adaptive step sizes, which is inherently faster than the DDPM sampler.