Text-to-3D Generation Quiz

Answers and explanations appear as you take the quiz.

1. What was the key innovation of DreamFusion (2022)?
   • It used Score Distillation Sampling to make 3D from a 2D diffusion model without 3D training data
   • It was the first to scan real objects with LiDAR
   • It trained directly on a massive labeled 3D mesh dataset
   • It generated 3D models instantly using a single forward pass
2. What is the 'Janus problem' in text-to-3D?
   • Models render too slowly to be useful
   • Generated objects show multiple faces or fronts from different angles
   • Textures appear washed out and gray
   • The 3D file format is incompatible with game engines
3. Which pair were OpenAI's early text-to-3D feed-forward models?
   • DALL-E and CLIP
   • Stable Diffusion and Midjourney
   • Point-E and Shap-E
   • DreamFusion and Magic3D
4. Why were early optimization-based methods like DreamFusion slow?
   • They had to download models from the cloud each time
   • They rendered in 8K resolution by default
   • They required manual sculpting between steps
   • They optimized a separate 3D representation per object over many rendering iterations
5. Which output format is NOT a typical 3D representation produced by these systems?
   • MP3 audio waveform
   • Polygon mesh
   • Radiance field (NeRF)
   • Point cloud
6. What does 'multi-view consistency' refer to as a challenge?
   • Making the file size consistent across exports
   • Ensuring the object looks coherent and correct from every camera angle
   • Keeping the text prompt under a word limit
   • Matching the color palette to a brand guide
7. What representation does Shap-E directly generate?
   • Only a fixed-resolution RGB image
   • A text transcript with speaker labels
   • Parameters of implicit functions that can render as textured meshes or neural radiance fields
   • A single unstructured audio waveform
8. How does the Point-E paper’s two-stage system produce a 3D point cloud?
   • A 3D artist manually models every prompt
   • A NeRF is optimized from scratch using hundreds of camera views
   • A classifier predicts a file extension without generating geometry
   • A text-to-image model first generates a synthetic view, then an image-conditioned diffusion model produces a point cloud