Null-Text Inversion Quiz

Answers and explanations appear as you take the quiz.

1. What does null-text inversion actually optimize?
   • The empty (null) text embedding used in classifier-free guidance
   • The weights of the entire diffusion U-Net
   • The pixels of the input image directly
   • The CLIP image encoder
2. Why does plain DDIM inversion become unreliable when used for editing?
   • It only works on synthetic images
   • Classifier-free guidance amplifies its reconstruction drift
   • It requires retraining the model
   • It cannot run in latent space
3. What is the main goal of inverting a real image before editing it?
   • To compress the image for storage
   • To remove the image's metadata
   • To find a noise/latent starting point that regenerates that exact image
   • To convert it to grayscale
4. Roughly how does null-text inversion handle the optimization across the diffusion process?
   • It optimizes a single shared embedding for the whole image
   • It skips optimization entirely
   • It optimizes only the final timestep
   • It optimizes a separate null embedding for each denoising timestep
5. A key downside of the original null-text inversion is that it is:
   • Per-image optimization, making it relatively slow
   • Unable to use text prompts
   • Only usable on 8x8 thumbnails
   • Incompatible with Stable Diffusion
6. What does Prompt-to-Prompt use to control edits while retaining parts of an image layout?
   • A trainable 3D Gaussian representation
   • Cross-attention maps associated with prompt tokens
   • A classifier that labels every pixel as a new object
   • A new text encoder trained for each edited image
7. What role does pivotal inversion play in null-text inversion?
   • It replaces the diffusion model with a newly trained classifier
   • It deletes the input image before reconstruction
   • It provides a per-timestep noise trajectory that anchors later optimization
   • It selects a new text prompt without using the image
8. During null-text optimization, which components does the method keep fixed?
   • Only the input image dimensions
   • The model weights and every unconditional embedding at initialization
   • The image caption and all diffusion timesteps
   • The pretrained model weights and the conditional text embedding