ESRGAN and GAN Super-Resolution Quiz

Answers and explanations appear as you take the quiz.

1. What does the 'GAN' in ESRGAN refer to?
   • Generative Adversarial Network
   • Gradient Accumulation Network
   • Gaussian Approximation Node
   • Global Attention Network
2. What building block does the ESRGAN generator primarily use?
   • Long Short-Term Memory cells
   • Residual-in-Residual Dense Blocks (RRDB)
   • Capsule routing layers
   • Transformer self-attention blocks
3. Why did ESRGAN's authors remove batch normalization from the generator?
   • It slowed training too much
   • It increased memory beyond limits
   • It tended to introduce visual artifacts
   • It was incompatible with GANs
4. What is the main trade-off ESRGAN makes compared to pixel-loss-only methods?
   • Speed over accuracy
   • Color accuracy over resolution
   • Lower memory over higher resolution
   • Perceptual realism over exact pixel fidelity
5. Where is ESRGAN's perceptual (VGG) loss measured?
   • On features before activation
   • On raw RGB pixels only
   • On the discriminator's final score
   • On Fourier-transformed inputs
6. What innovation did ESRGAN's discriminator introduce?
   • A purely convolutional autoencoder
   • A relativistic discriminator comparing real vs. fake realism
   • A reinforcement-learning reward model
   • A frozen pretrained classifier
7. Which competition did the ESRGAN paper report winning?
   • The ImageNet classification challenge
   • The Vesuvius Challenge
   • The PIRM 2018 perceptual super-resolution challenge
   • The COCO image-captioning challenge
8. Why should generated detail from ESRGAN not be treated as recovered photographic evidence?
   • The model records the original sensor data exactly
   • GANs can only enlarge images without changing pixels
   • The output is guaranteed to match the lost high-resolution image
   • The model can synthesize plausible textures that were not present in the low-resolution input