Progressive Growing of GANs Quiz

Answers and explanations appear as you take the quiz.

1. How does a progressively grown GAN begin training?
   • At the full target resolution immediately
   • At a very low resolution such as 4x4 pixels
   • By pretraining only the discriminator
   • With a text prompt
2. What is the main benefit of growing resolution gradually?
   • It makes training more stable and faster than training at high resolution directly
   • It removes the need for a generator
   • It eliminates the dataset requirement
   • It guarantees zero artifacts
3. When a new higher-resolution layer is added, how is it introduced?
   • Instantly at full strength
   • Only in the discriminator, never the generator
   • Smoothly faded in with a blending weight (alpha) that ramps from 0 to 1
   • By retraining the network from scratch
4. Why are layers added to both the generator and discriminator?
   • Because the generator alone cannot produce color
   • To reduce the dataset size
   • Only for visualization purposes
   • To keep the two networks symmetric and balanced as resolution increases
5. What is the purpose of the minibatch standard deviation layer introduced in ProGAN?
   • To upscale the final image
   • To compress the model
   • To increase output variation and discourage mode collapse
   • To label the training images
6. What later development reduced the need for explicit progressive growing?
   • The invention of the discriminator
   • StyleGAN2 showing a fixed architecture with skip/residual connections could match quality
   • Switching to grayscale images
   • Removing the latent vector entirely
7. At alpha equal to zero during a ProGAN generator fade-in, which path supplies the blended output?
   • Only the newly added block at full strength
   • An unrelated pretrained text encoder
   • A random photograph taken from the test set
   • The previous lower-resolution path after upsampling
8. What is the original ProGAN equalized-learning-rate method intended to balance?
   • The effective scale of learning across layers through dynamic weight scaling
   • The number of captions assigned to each face image
   • The timestamp of every generated training sample
   • The count of output file formats across batches