Imagen Text-to-Image Quiz

Answers and explanations appear as you take the quiz.

1. In the original Imagen study's scaling comparison, which change produced larger image-quality and text-alignment gains?
   • Discarding the language encoder entirely
   • Increasing the pretrained text encoder's size rather than the diffusion model's size
   • Replacing diffusion with nearest-neighbor image lookup
   • Reducing every prompt to a single word
2. Which text encoder does Imagen rely on?
   • A convolutional image net
   • A small LSTM
   • A hand-written rule set
   • A frozen T5-XXL language model
3. How does Imagen reach high resolution?
   • Cascaded diffusion with super-resolution upscaling stages
   • A single giant generation step
   • By cropping a large stock photo
   • By interpolating two images
4. What problem did original Imagen's dynamic thresholding address under strong guidance?
   • A lack of text tokens in the encoder vocabulary
   • The need to compress the final PNG file
   • Saturation of predicted pixel values during sampling
   • A missing training label for every object
5. What is DrawBench?
   • A drawing tablet
   • A GPU brand
   • A dataset of cat photos
   • A benchmark of challenging prompts for evaluating text-to-image models
6. Why is the T5-XXL encoder kept frozen in Imagen?
   • It is broken and cannot train
   • Its rich text understanding from pretraining transfers without further tuning
   • To save disk space only
   • Because text is irrelevant
7. What resolution did original Imagen's base diffusion model generate before its super-resolution stages?
   • 256 by 256 pixels as the first stage
   • 1024 by 1024 pixels directly
   • 64 by 64 pixels
   • A single pixel that was copied across the final image
8. Why did the Imagen researchers add Gaussian noise to low-resolution conditioning images during super-resolution training?
   • To make the upsamplers more robust to imperfect inputs from earlier stages
   • To remove text conditioning from the entire cascade
   • To guarantee that all training images had the same caption
   • To replace the upsamplers with deterministic resizing