StyleGAN Architecture Quiz

Answers and explanations appear as you take the quiz.

1. What is the purpose of StyleGAN's mapping network?
   • To classify whether an image is real or fake
   • To upsample images to higher resolution
   • To compress the final image for storage
   • To transform the latent z into a more disentangled intermediate space W
2. In the original StyleGAN, how is style injected into the feature maps?
   • By concatenating the latent vector to each pixel
   • Through Adaptive Instance Normalization (AdaIN)
   • Via a final softmax layer
   • Using max pooling
3. What does the synthesis network start from instead of the latent vector?
   • A learned constant tensor
   • A photograph of a real face
   • Pure white noise the same size as the output
   • An edge map
4. What does adjusting styles at the coarse (low-resolution) layers primarily control?
   • Fine skin texture and freckles
   • Color of individual pixels only
   • High-level attributes like pose and face shape
   • Image file format
5. What problem did StyleGAN2 fix relative to the original?
   • It added text-to-image prompting
   • It removed characteristic blob-like artifacts caused by AdaIN
   • It eliminated the need for a discriminator
   • It made training fully unsupervised for the first time
6. What is 'style mixing' in StyleGAN?
   • Blending two datasets before training
   • Averaging the discriminator's outputs
   • Mixing audio with the generated image
   • Using different w vectors at different layers to combine attributes from two images
7. What role do separate per-layer noise inputs play in original StyleGAN synthesis?
   • They replace the mapping network's entire style representation
   • They determine the file's compression format
   • They supply stochastic fine detail, such as varying hair strands or freckles
   • They provide the discriminator's real-image labels
8. What visual behavior did StyleGAN3's alias-free design target?
   • Texture sticking and aliasing that make details move unnaturally as image content changes
   • The absence of a text tokenizer in original StyleGAN
   • A requirement that every output be a grayscale photograph
   • The need to store generated images as audio waveforms