Custom Diffusion Multi-Concept Tuning Quiz

Answers and explanations appear as you take the quiz.

1. Which part of the diffusion model does Custom Diffusion primarily fine-tune?
   • The key and value projections in cross-attention layers
   • Every weight in the U-Net
   • The VAE decoder only
   • The text tokenizer
2. What is Custom Diffusion most notable for compared to single-concept methods?
   • Generating video instead of images
   • Composing multiple newly learned concepts in one scene
   • Requiring no images at all
   • Running only on CPUs
3. What minimum number of training images did the Custom Diffusion paper report using in some experiments?
   • As few as four images
   • At least 500 images
   • Exactly one image in every experiment
   • The paper specifies no concept images
4. Why does Custom Diffusion use a set of regularization images from the concept's broader category?
   • To speed up the VAE
   • To increase image resolution
   • To remove watermarks
   • To prevent overfitting and language drift on the concept word
5. How can two separately trained Custom Diffusion concepts be used together?
   • By merging them via a constrained optimization over the tuned weights
   • They cannot ever be combined
   • By averaging the input images
   • By retraining the full model from scratch
6. What proportion of model weights did the Custom Diffusion paper report storing for its tuned subset?
   • About 3% of the model weights
   • All model weights plus a new model copy
   • No trained parameters at all
   • A fixed 50% of the model weights
7. Besides merging separately tuned models, how does Custom Diffusion support learning multiple concepts?
   • It requires a separate foundation model for every prompt
   • It replaces all cross-attention with a fixed lookup table
   • It can train for multiple concepts jointly
   • It can only learn one concept per model
8. In a cross-attention layer, what do the tuned text-side key and value projections help associate?
   • Image pixels with internet search results
   • The diffusion timestep with the model's source code
   • Audio frequencies with a document's page number
   • Prompt concepts with features in the image latent representation