GLIDE Diffusion Model Quiz

Answers and explanations appear as you take the quiz.

1. Which guidance method did GLIDE find produced better, more prompt-faithful images?
   • Classifier-free guidance
   • CLIP guidance
   • No guidance
   • Manual editing
2. What size did the published full GLIDE base diffusion model have, before counting its upsampler?
   • About 3.5 thousand parameters
   • About 350 billion parameters
   • About 3.5 billion parameters
   • No trainable parameters
3. How does classifier-free guidance work during training?
   • The model only ever sees the correct prompt
   • A separate image classifier is trained alongside
   • Prompts are translated to numbers randomly
   • The model sometimes sees the real prompt and sometimes a blank one
4. Which later system describes its diffusion decoder as modifying the GLIDE architecture in the DALL-E 2 research paper?
   • The original text-only GPT-2 model
   • The unCLIP system associated with DALL-E 2
   • The Whisper speech recognizer
   • The CLIP contrastive encoder alone
5. Why did OpenAI release only a smaller, filtered version of GLIDE publicly?
   • The full model was lost
   • It ran only on supercomputers
   • Concerns about misuse of a powerful image generator
   • Licensing from a third party
6. What practical editing feature did GLIDE support?
   • Text-driven inpainting
   • Audio transcription
   • Spreadsheet formulas
   • Video rendering
7. How did GLIDE's upsampler complement its base text-conditioned diffusion model?
   • It replaced the text prompt with an audio transcript
   • It increased the resolution of the image produced by the base model
   • It converted the image into the training tokenizer's vocabulary
   • It supplied manual labels for every generated object
8. At GLIDE sampling time, how does classifier-free guidance strengthen prompt conditioning?
   • It retrains the model weights after every generated pixel
   • It requires a separate class-label classifier to score every step
   • It replaces the conditional prediction with the unconditional one
   • It combines conditional and unconditional predictions, extrapolating toward the conditional one