Викторина по моделям диффузии

Ответы и пояснения появляются по мере прохождения теста.

1. How does classic DDPM sampling produce a generated image from an initial noise sample?
   • By copying a required identical training image
   • By assigning the output filename as its pixel values
   • By performing no model computation
   • By following successive learned reverse-process steps
2. In the common noise-prediction DDPM parameterization, what training target does the network estimate?
   • The market price of the image
   • The Gaussian noise added to the training example
   • The image's copyright owner
   • A required final file extension
3. In the text-conditioned latent diffusion formulation, how can prompt embeddings influence denoising?
   • Through cross-attention conditioning
   • Only by changing the file extension
   • By removing every learned layer
   • By guaranteeing an exact training-image copy
4. Why does latent diffusion perform denoising in a compressed representation?
   • To guarantee identical reconstruction to every pixel-space model
   • To eliminate all neural computation
   • To reduce the dimensionality of the heavy denoising computation
   • To make every output necessarily monochrome
5. Which component is not required by the classic DDPM denoising training objective?
   • Training examples
   • An adversarial discriminator trained to distinguish real from generated samples
   • A defined corruption process
   • A learned reverse-process model
6. What does progressive distillation of a diffusion sampler aim to reduce?
   • The need to evaluate generation quality
   • Every condition supplied to the model
   • The number of pixels in every possible output by definition
   • The number of sampling steps or model evaluations
7. After denoising in a latent diffusion image model, which component reconstructs image pixels from the latent representation?
   • A filename formatter
   • The original training-data shuffle
   • The autoencoder decoder
   • An adversarial label threshold in every model
8. What does a DDPM forward noise schedule specify across successive steps?
   • How much noise variance is introduced
   • A guarantee that every sample is factually true
   • The copyright license for the training data
   • The alphabetical ordering of prompt words