Textual Inversion Quiz

Answers and explanations appear as you take the quiz.

1. What kinds of concepts did the original textual-inversion paper personalize?
   • User-provided objects or visual styles
   • Only database schemas
   • Only speech transcripts
   • Only camera calibration parameters
2. Roughly how many example images does Textual Inversion typically need?
   • A single image only
   • About 3 to 5 images
   • Thousands of images
   • No images at all
3. Why are Textual Inversion files so small (often a few kilobytes)?
   • They compress the whole model
   • They are just text prompts
   • They store only one learned vector
   • They reference an online database
4. What stays unchanged during Textual Inversion training?
   • The example images
   • The new pseudo-word
   • The text prompt format
   • All of the model's existing weights
5. How do you use a learned concept after Textual Inversion?
   • Type its new pseudo-word in a prompt
   • Retrain the model each time
   • Upload the photos again
   • Edit the model's source code
6. Which loss is used to optimize the new embedding?
   • A classification accuracy loss
   • The diffusion denoising loss on the example images
   • A speech-recognition loss
   • No loss; it's manual tuning
7. How does textual inversion use the learned token at generation time?
   • It replaces the image decoder
   • It directly edits a generated image without sampling
   • It inserts the token into a text prompt so the frozen text-to-image model can condition on it
   • It changes the model’s tokenizer vocabulary for every user
8. What is a key limitation of a concept learned from only a few example images?
   • It always increases the model’s image resolution
   • It automatically learns every view and lighting condition
   • It changes all existing model concepts equally
   • It may capture incidental features of those examples and generalize imperfectly