Autoregressive Image Generation Quiz

Answers and explanations appear as you take the quiz.

1. What does 'autoregressive' mean in the context of image generation?
   • Each new element is predicted conditioned on previously generated elements
   • Each new element is predicted using only the text prompt
   • The whole image is denoised at once from random noise
   • Pixels are generated completely independently of one another
2. How do modern autoregressive image models like DALL-E 1 typically represent an image before predicting it?
   • As a single continuous vector
   • As a fixed grid of discrete tokens from a learned codebook
   • As raw 24-bit RGB pixels only
   • As a list of vector graphics shapes
3. Which early models generated images one raw pixel at a time?
   • GANs and VAEs
   • Stable Diffusion and Imagen
   • PixelRNN and PixelCNN
   • CLIP and BLIP
4. What mechanism ensures a Transformer only attends to earlier tokens during autoregressive generation?
   • Dropout
   • Batch normalization
   • Skip connections
   • Causal (masked) self-attention
5. What is the main practical drawback of autoregressive image sampling?
   • Sampling is sequential and therefore slow
   • It cannot be conditioned on text
   • It cannot model fine detail
   • It requires no training data
6. Which technique speeds up token-based image generation by predicting many tokens in parallel?
   • Teacher forcing
   • Masked parallel decoding (e.g., MaskGIT/Muse)
   • Gradient clipping
   • One-hot encoding
7. How does the original DALL·E model use text and image tokens?
   • Separate unconnected classifiers process text and pixels
   • A diffusion decoder predicts all pixels simultaneously
   • A single autoregressive Transformer models the sequence conditioned on text and image tokens
   • The model uses only a text encoder and no image-token sequence
8. What scale of image-text pairs did the original DALL·E paper report collecting?
   • About 250 pairs
   • About 25 thousand pairs
   • About 25 billion pairs
   • About 250 million image-text pairs