DreamBooth Quiz

Answers and explanations appear as you take the quiz.

1. What does DreamBooth modify that Textual Inversion does not?
   • The model's actual weights
   • Only the prompt text
   • The image file format
   • The user's GPU drivers
2. What is the purpose of DreamBooth's prior-preservation loss?
   • To speed up image rendering
   • To prevent overfitting and keep the broader class concept
   • To shrink the file size
   • To remove backgrounds
3. How is a subject typically referenced in DreamBooth training prompts?
   • By its full real name
   • By a numeric ID only
   • By a rare token plus a class word, like 'sks dog'
   • By an emoji
4. Roughly how many subject images does DreamBooth need?
   • Thousands
   • Exactly one
   • One hundred
   • Around 3 to 5
5. What is a common trade-off of using DreamBooth?
   • Larger storage and more compute than lightweight methods
   • Lower image quality than embeddings
   • It only works on text, not images
   • It needs no training at all
6. Which combination is now widely used to make DreamBooth lighter?
   • DreamBooth with OCR
   • DreamBooth with LoRA
   • DreamBooth with GPS
   • DreamBooth with audio models
7. What subject variations did the DreamBooth paper demonstrate in generated images?
   • Only the exact camera angle from the training photos
   • Only the original background pixels
   • The subject could appear in novel scenes, poses, views and lighting
   • Different subjects with no relation to the reference
8. What kind of model parameters does LoRA adapt?
   • Only the image file metadata
   • A new pixel palette for every output
   • The dataset labels without changing model computation
   • Low-rank update matrices added to selected weight transformations