GLIGEN Grounded Generation Quiz

Answers and explanations appear as you take the quiz.

1. What problem does GLIGEN primarily solve?
   • Lack of precise spatial control over where objects appear
   • Slow generation speed
   • Inability to generate color images
   • Missing audio in generated videos
2. How does GLIGEN preserve the knowledge of the pretrained diffusion model?
   • It retrains all weights from scratch
   • It freezes the original weights and adds new trainable gated self-attention layers
   • It deletes the text encoder
   • It uses a GAN instead of diffusion
3. What is a typical grounding input GLIGEN accepts?
   • A musical score
   • A list of hyperparameters
   • A bounding box paired with a text or image entity
   • A spreadsheet of pixel values
4. Why is the gate in GLIGEN's new attention layers initialized to zero?
   • To make images grayscale at first
   • To disable the text prompt
   • To reduce the image resolution
   • So the base model's behavior is preserved early in training and control is added gradually
5. What does 'open-world' grounded generation mean in GLIGEN?
   • It can ground arbitrary described concepts and layouts not seen during grounding training
   • It only works on indoor scenes
   • It requires a 3D world model
   • It generates only outdoor landscapes
6. Where are GLIGEN's new gated self-attention layers inserted in the U-Net?
   • Only at the output VAE
   • Between the existing self-attention and cross-attention blocks
   • Inside the text encoder only
   • Before the noise scheduler
7. What does a GLIGEN grounding token combine?
   • A denoising timestep with a VAE decoder output
   • A color histogram with the final image resolution
   • A text or image embedding with spatial information such as a bounding box
   • A training dataset name with a classifier label
8. Besides bounding boxes, which inputs can GLIGEN use to ground objects?
   • Only unstructured audio waveforms
   • A crop-size label without any object information
   • A list of model checkpoint filenames
   • Keypoints or reference images