DreamFusion and Score Distillation Sampling Quiz

Answers and explanations appear as you take the quiz.

1. What 3D representation does DreamFusion optimize?
   • A NeRF (Neural Radiance Field)
   • A polygon mesh
   • A point cloud from a depth sensor
   • A voxel grid from CT data
2. How much 3D training data does DreamFusion require?
   • A few hundred scans
   • None
   • Millions of CAD models
   • One reference object per prompt
3. What is the computational shortcut in Score Distillation Sampling?
   • It trains a new diffusion model every step
   • It uses no gradients at all
   • It uses a stop-gradient score signal instead of backpropagating through the diffusion model’s full Jacobian
   • It renders only one fixed viewpoint
4. Why did DreamFusion use high classifier-free guidance in its experiments?
   • To reduce GPU memory
   • To make outputs grayscale
   • To guarantee accurate geometry
   • To counter the oversmoothing tendency of SDS and encourage sharper prompt-conditioned results
5. Which 2D model did the original DreamFusion use as its frozen prior?
   • Google's Imagen
   • Stable Diffusion 1.5
   • DALL-E 2
   • Midjourney
6. What artifact can arise when a 2D image prior guides a 3D scene across views?
   • A guaranteed watertight mesh
   • Duplicated faces or features on different sides of the object
   • A single grayscale output
   • A 3D object with no rendered views
7. Why does DreamFusion render the NeRF from sampled camera angles during optimization?
   • To avoid using the NeRF representation
   • To replace the text prompt with camera metadata
   • To let the 2D prior provide guidance from multiple views of the 3D scene
   • To create an audio soundtrack
8. What does an optimized NeRF allow a user to do with the generated scene?
   • Recover the original 3D training scan
   • Guarantee physically correct materials
   • Generate only the original training image
   • Render it from new viewpoints and relight it