Latent Diffusion Models Quiz

Answers and explanations appear as you take the quiz.

1. What is the main innovation of latent diffusion models compared to pixel-space diffusion?
   • They run the diffusion process in a compressed latent space instead of on raw pixels
   • They eliminate the need for any neural network
   • They only work on black-and-white images
   • They require no training data at all
2. In latent diffusion, what does the first-stage autoencoder do?
   • Predicts text tokens without processing images
   • Encodes an image into a latent representation and decodes latents back into images
   • Replaces the diffusion model at every denoising step
   • Classifies a prompt as safe or unsafe
3. How is text typically injected into the denoising U-Net in latent diffusion?
   • By concatenating the raw text string to the image
   • By manually editing pixels
   • Through cross-attention layers using text embeddings
   • Text cannot be used at all
4. Why can denoising in a latent diffusion model reduce computation compared with pixel-space diffusion?
   • The model skips all image decoding
   • The diffusion network uses no neural layers
   • The model removes every denoising step
   • The denoiser operates on a smaller spatial latent grid, though decoding is still needed at the end
5. Which first-stage autoencoder regularization variants does the original latent diffusion paper investigate?
   • KL-regularized and vector-quantized autoencoders
   • Only a classifier-free guidance model
   • Only a causal language model
   • A decoder that has no learned parameters
6. What does the diffusion model learn to do during training?
   • To classify images into categories
   • To reverse a gradual noising process and denoise step by step
   • To compress text into audio
   • To translate languages
7. In the paper’s f=8 autoencoder configuration, what does the downsampling factor describe?
   • The model uses eight separate text encoders
   • Every input is reduced to exactly eight pixels
   • Each spatial dimension of the latent grid is reduced by a factor of eight
   • The diffusion model performs eight total training updates
8. How does the latent diffusion paper inject spatial conditions such as semantic layouts into the denoiser?
   • It converts the layout into an English sentence only
   • It removes spatial coordinates before denoising
   • It uses the condition only after decoding to pixels
   • It concatenates the spatial condition with the latent representation along channels