Google Imagen Quiz

Answers and explanations appear as you take the quiz.

1. What type of generative model is Google Imagen built on?
   • A diffusion model
   • A generative adversarial network (GAN)
   • A variational autoencoder only
   • A rule-based renderer
2. A key finding from the original Imagen paper was that scaling which component most improved results?
   • The image diffusion network
   • The frozen text encoder (language model)
   • The training batch size
   • The number of color channels
3. Which improvement did Google announce for Imagen 4 in the Gemini app?
   • Saving files as PNG
   • Generating audio
   • Improved text rendering in generated images
   • Compressing images
4. Imagen's original architecture used a cascade that started by generating an image at what resolution?
   • 1024x1024
   • 512x512
   • 4096x4096
   • 64x64
5. What did Google DeepMind say SynthID does for images generated with Imagen on Vertex AI?
   • Embeds an imperceptible watermark in image pixels to help identify likely Imagen output
   • Converts all images to a new file format
   • Provides a GPU chip for image training
   • Translates the prompt language
6. Through which Google platform can developers access Imagen for their own applications?
   • Google Sheets
   • Vertex AI
   • Google Maps
   • Chrome DevTools
7. What issue does dynamic thresholding address in Imagen's sampling process?
   • The model having too few text tokens in its prompt
   • The training set containing too many image resolutions
   • Pixel values becoming oversaturated when classifier-free guidance is high
   • The tokenizer splitting a word into subwords
8. What does classifier-free guidance do during Imagen sampling?
   • Selects a class label after the image is complete
   • Upscales the image without conditioning on its prompt
   • Replaces the noise-prediction model with a fixed image lookup
   • Combines conditional and unconditioned predictions to steer generation toward the prompt