Muse Masked Generative Imaging Quiz

Answers and explanations appear as you take the quiz.

1. What does Muse predict during generation instead of denoising pixels?
   • Raw RGB pixel values directly
   • Masked discrete image tokens from a VQGAN codebook
   • Vector strokes like a vector graphics program
   • Depth maps that are later colored in
2. What feature contributes to Muse's reported sampling efficiency relative to the evaluated sequential image-token baseline?
   • Generating only monochrome outputs
   • Removing all text-encoder computation
   • Guaranteeing completion of every image in a single pass
   • Predicting multiple masked image tokens in each decoding round
3. How does original Muse condition image prediction on its text prompt?
   • With embeddings from a pretrained frozen T5-XXL encoder
   • With an untrained word-frequency table
   • By reading only the first character
   • By turning off text processing during inference
4. What is the role of confidence-based remasking in Muse?
   • It deletes low-quality training images
   • It compresses the final image file
   • It keeps confident token predictions and re-masks uncertain ones for the next round
   • It rates the user's prompt quality
5. How does Muse's two-stage token pipeline produce a higher-resolution image representation?
   • It enlarges the output only with a blur filter
   • It predicts all resolutions with no additional model
   • It discards the text condition before the second stage
   • A super-resolution token model uses the base representation to predict a higher-resolution token grid
6. What type of task is Muse naturally suited for because it predicts masked regions?
   • Audio transcription
   • Inpainting and outpainting
   • Spreadsheet calculation
   • Video compression only
7. After Muse predicts its image-token grid, which component reconstructs pixels?
   • The text prompt's keyboard input handler
   • An AIS transponder
   • The corresponding VQGAN tokenizer decoder
   • The training-caption file name
8. Why does a quoted Muse image-model parameter count omit some parameters used by the full generation system?
   • The frozen T5-XXL text encoder is an additional component
   • Every frozen model has zero stored parameters
   • The model count measures only PNG file bytes
   • The text encoder becomes the image tokenizer at runtime