Fréchet Inception Distance Quiz

Answers and explanations appear as you take the quiz.

1. What does a lower FID score indicate?
   • The generated images' statistics are closer to those of real images
   • The generated images are less realistic
   • The model trained for fewer epochs
   • The images have higher resolution
2. Which pretrained network does standard FID use to extract image features?
   • ResNet-50
   • Inception-v3
   • VGG-16
   • BERT
3. How does FID summarize each set of image features before comparing them?
   • As a single average pixel value
   • As a sorted list of file names
   • As a multivariate Gaussian with a mean vector and covariance matrix
   • As a histogram of color channels only
4. What key shortcoming of the Inception Score did FID address?
   • Inception Score was too slow to compute
   • Inception Score only worked on text
   • Inception Score required no neural network
   • Inception Score never compared generated images to real data
5. When a generator produces less varied Inception features than the real images, how can that affect FID?
   • A covariance mismatch can contribute to a higher FID
   • FID must become zero
   • FID ignores feature diversity
   • FID automatically identifies the exact cause of the collapse
6. In the Diffusion Models Beat GANs comparison, how was FID evaluated for ImageNet?
   • With 500 generated samples
   • With 50,000 generated samples
   • With 5,000 generated samples
   • With 5 million generated samples
7. Why can a small evaluation set make an FID estimate unreliable?
   • FID requires every image to have a human label
   • The Inception network stops extracting features on small sets
   • Finite-sample bias can distort the estimated distance between feature distributions
   • Small sets force all covariance values to zero
8. What does Kernel Inception Distance (KID) offer as an alternative in the guide?
   • A method for training the Inception-v3 network
   • A way to generate image captions
   • A measure of video frame rate
   • A metric used for evaluation when sample counts are smaller