Quiz sur le mélange latent et l'interpolation d'images

Les réponses et les explications apparaissent au fur et à mesure que vous répondez au quiz.

1. When a pretrained generator decodes an interpolated latent code, what does the result represent?
   • A model-specific generated intermediate sample whose quality depends on its latent space
   • A guaranteed pixel-perfect average of the two source images
   • A copy of one endpoint image with no new decoding
   • A universal semantic blend that works identically across generators
2. When averaging two unrelated images directly in pixel space, what visual effect can result?
   • Une nouvelle scène pointue
   • Un look double exposition transparent et fantôme
   • Une morphologie parfaite
   • Une image de plus haute résolution
3. What does the Sampling Generative Networks paper report spherical interpolation can do?
   • Guarantee that every path follows the data manifold
   • Make any two unrelated images semantically compatible
   • Help avoid diverging from the model’s prior distribution and produce sharper samples in its examples
   • Remove the need to decode intermediate latent codes
4. What latent-space visualization techniques does the cited paper introduce?
   • A system for synchronizing captions to music video
   • A method that estimates image-file metadata
   • A tool for converting latent vectors into spoken audio
   • J-Diagrams and MINE grids for visualizing manifolds from analogies and nearest neighbors
5. What does the 2021 latent-interpolation paper measure to compare a straight line with a shorter path?
   • Relative path-length improvement under the generator’s pull-back metric
   • The number of pixels shared by both endpoint images
   • The number of interpolation frames rendered per second
   • The text-prompt length of the decoded samples
6. What does the 2021 study say about latent-space geometry in deep generative models?
   • It is often model-dependent and may be non-Euclidean, so geometry should be evaluated
   • It is always a Euclidean space with identical geometry across models
   • It is always a hypersphere where slerp guarantees quality
   • It has no relation to the generator’s mapping
7. After interpolating two image latents, what step produces a visible in-between sample?
   • Sort the latent coordinates alphabetically
   • Replace the latent with a text caption only
   • Decode the intermediate latent with the generative model
   • Average the two images' file sizes
8. In a weighted latent blend, what does the interpolation coefficient control?
   • The number of training images in the dataset
   • The resolution of the decoder's output layer
   • Whether the model uses labels during pretraining
   • The relative contribution of the two endpoint latent codes