GigaGAN Scaled Generators Quiz

Answers and explanations appear as you take the quiz.

1. What total parameter count did the paper report for one high-capacity GigaGAN framework configuration?
   • 1.0 billion total across the 652.5M base generator and 359.1M upsampler
   • 70 million total across a 64M generator and 6M discriminator
   • 1.0 billion in the base generator alone, with no upsampler
   • 3.5 billion total across two text encoders
2. What is a major speed advantage of GigaGAN over diffusion models?
   • It needs no GPU
   • It generates an image in a single forward pass instead of many denoising steps
   • It trains in one second
   • It uses no neural network
3. What does GigaGAN's 'sample-adaptive kernel selection' do?
   • Deletes unused pixels
   • Compresses the image to JPEG
   • Blends a bank of convolution filters using weights derived from the text prompt
   • Picks the GPU automatically
4. How does GigaGAN incorporate text information into image generation?
   • By cropping the prompt
   • It ignores text entirely
   • By printing the text on the image
   • Through cross-attention to text embeddings and matching CLIP features
5. Which extra capability does GigaGAN provide beyond base generation?
   • A separate GAN upsampler that can reach 4K resolution
   • Audio synthesis
   • Speech recognition
   • Database querying
6. Why is GigaGAN's latent space useful for editing tools?
   • It cannot be edited
   • It supports smooth interpolation and style mixing for controllable changes
   • It only stores text
   • It deletes the image after one use
7. How does GigaGAN’s image discriminator make predictions across resolutions?
   • It scores only the text prompt and ignores generated images
   • It predicts one label at the final resolution only
   • It receives an image pyramid and predicts real or fake at multiple input and downsampling scales
   • It compares generated images with a fixed set of training captions only
8. What does GigaGAN’s base generator produce for later multi-scale processing?
   • A single class label for the entire image
   • A sequence of text tokens for an autoregressive decoder
   • Only the final upsampled image, with no intermediate scales
   • An image pyramid with outputs at several resolutions