Virtual Try-On with Diffusion Models Quiz

Answers and explanations appear as you take the quiz.

1. What are the two competing goals every virtual try-on system must balance?
   • Maximizing resolution and minimizing file size
   • Keeping the person unchanged and faithfully reproducing the garment
   • Matching the garment's price and the shopper's budget
   • Generating new faces and new backgrounds
2. How did early methods such as VITON and CP-VTON mainly work?
   • They warped the flat garment to the pose and then blended it into the photo
   • They built a full 3D physics simulation of the fabric
   • They searched a database for a real photo of the person wearing the item
   • They used a large language model to describe the outfit
3. Why did 2D warping struggle with crossed arms or large pose changes?
   • Warping is too slow to run on phones
   • A warp cannot invent garment parts that were never visible in the source image
   • Warping only works on black and white images
   • Warping requires the shopper's exact measurements
4. In a diffusion try-on model, what does the model do to the clothing region?
   • Copies it from the garment image pixel by pixel
   • Blurs it to hide artifacts
   • Regenerates it from noise while conditioned on the garment, pose and masked photo
   • Leaves it untouched and changes only the background
5. What design feature did Google's TryOnDiffusion use to transfer the garment?
   • Two parallel networks exchanging information through cross-attention
   • A separate text prompt describing each garment
   • A physical mannequin scanned in 3D
   • A rule-based system of fixed templates
6. What is the most important thing a try-on image does NOT tell you?
   • The color of the garment
   • The general style of the garment
   • Whether the garment will actually fit your body
   • What the garment looks like from the front
7. Why do many systems extract garment features with a reference network instead of a single CLIP embedding?
   • CLIP embeddings are illegal to use commercially
   • A reference network preserves fine details such as logos and prints better
   • A reference network removes the need for pose maps
   • CLIP cannot process color images
8. Why are metrics like FID insufficient for judging a try-on system?
   • They only work on text
   • They cannot be computed on images
   • They measure overall realism but say little about fit or fairness across body types
   • They always rate diffusion models as worse