Real-ESRGAN Practical Restoration Quiz

Answers and explanations appear as you take the quiz.

1. What key weakness of the original ESRGAN did Real-ESRGAN address?
   • It was trained only on clean bicubic downscaling
   • It only worked on grayscale images
   • It could not run on GPUs
   • It required paired real photos
2. What is Real-ESRGAN's 'high-order degradation' model?
   • A higher-resolution generator
   • A chain of multiple random degradations applied in sequence
   • A new attention mechanism
   • A larger discriminator
3. What discriminator architecture does Real-ESRGAN use?
   • A single linear classifier
   • A recurrent network
   • A U-Net with spectral normalization
   • A frozen VGG
4. Which artifacts do the 'sinc' filters in Real-ESRGAN reproduce?
   • Color banding
   • Lens flare
   • Chromatic aberration
   • Ringing and overshoot
5. What backbone does the Real-ESRGAN generator retain from ESRGAN?
   • The RRDB blocks
   • A transformer encoder
   • A diffusion U-Net
   • An LSTM stack
6. In GFPGAN’s documented inference command, which background upsampler is set by default?
   • Real-ESRGAN
   • ESRGAN only
   • A fixed bicubic resize
   • No background upsampler is supported
7. What kind of images is Real-ESRGAN’s anime checkpoint intended to restore?
   • Only medical CT scans
   • Only satellite photographs
   • Anime illustrations
   • Only grayscale handwritten documents
8. Why can tiled inference help when restoring a very large image?
   • It guarantees that every tile contains the same details
   • It replaces super-resolution with image compression
   • It prevents the model from seeing any neighboring pixels
   • It processes smaller patches to reduce peak memory use