Zero-1-to-3 Novel View Diffusion Quiz

Answers and explanations appear as you take the quiz.

1. What is the core input Zero-1-to-3 needs in addition to a single image to generate a new view?
   • A relative camera transform (rotation and translation)
   • A depth map of the scene
   • A second photo from another angle
   • A hand-drawn sketch of the target view
2. Zero-1-to-3 is built by fine-tuning which kind of model?
   • A graph neural network
   • A 2D image diffusion model (Stable Diffusion)
   • A recurrent sequence model
   • A classical structure-from-motion solver
3. Why can a 2D diffusion model perform zero-shot novel view synthesis at all?
   • Because it stores explicit 3D meshes internally
   • Because it uses LiDAR data during inference
   • Because it has learned implicit geometric and physical priors from large-scale image training
   • Because it memorizes every training image exactly
4. What dataset of 3D objects was central to training Zero-1-to-3?
   • ImageNet
   • COCO
   • KITTI driving scenes
   • Objaverse (synthetic rendered objects)
5. How is the source image's fine detail preserved in the generation?
   • By channel-concatenating the input image to the noisy latent
   • By discarding it after computing a caption
   • By using only its file name as a label
   • By converting it to grayscale first
6. What is a common downstream use of the views Zero-1-to-3 produces?
   • Audio transcription
   • Reconstructing a 3D model or mesh of the object
   • Spam email filtering
   • Stock price prediction
7. Which values describe the relative camera pose in Zero-1-to-3's conditioning?
   • Image brightness, contrast, and saturation
   • Object color, texture, and material
   • Azimuth, elevation, and radius
   • Frame rate, shutter speed, and aperture
8. Which input representation does Zero-1-to-3 use to steer cross-attention with image semantics?
   • A depth map produced by a stereo camera pair
   • A mesh extracted from the final generated view
   • A class label that replaces the source image
   • A CLIP embedding combined with the relative camera pose