VQGAN and Codebook Image Synthesis Quiz

Answers and explanations appear as you take the quiz.

1. What does VQGAN use to represent an image as discrete tokens?
   • A learned codebook of vectors that encoder outputs are snapped to
   • A list of file names
   • Raw RGB pixels only
   • A single number per image
2. What did the original VQGAN paper report gaining from adversarial and perceptual training?
   • To guarantee every VQVAE reconstruction is blurry
   • To preserve perceptual quality at increased compression in the original paper’s evaluated settings
   • To eliminate the need for a learned codebook
   • To make all images use the same token grid regardless of dataset
3. Once an image is a sequence of tokens, what model generates new images?
   • A spreadsheet
   • A calculator
   • An autoregressive transformer predicting tokens like words
   • A printer driver
4. What technique lets gradients flow through the non-differentiable quantization step?
   • Dropout
   • Batch normalization
   • Random cropping
   • The straight-through estimator
5. What famous AI art trend did VQGAN help create in 2021?
   • VQGAN+CLIP generative art
   • 3D printing
   • Voice cloning
   • Spreadsheet automation
6. What is a known challenge with VQGAN-style codebooks?
   • They cannot store any vectors
   • Codebook collapse, where many codes go unused
   • They require no training
   • They only work on audio
7. Why does VQGAN use a compact latent token grid for high-resolution images?
   • It removes the need for an image decoder
   • It makes every codebook entry a full-resolution image
   • It shortens the transformer sequence so modeling image composition is more tractable
   • It converts the transformer into a pixel-level convolution
8. Which loss does the VQGAN guide pair with adversarial training to improve reconstruction quality?
   • A language-model next-token loss
   • A contrastive audio-text loss
   • A reinforcement-learning reward from human rankings
   • Perceptual loss