Video Diffusion Models Quiz

Answers and explanations appear as you take the quiz.

1. What is the fundamental idea behind a diffusion model?
   • Learning to reverse a process that gradually adds noise, denoising step by step
   • Copying frames from existing videos
   • Compressing audio into images
   • Encrypting data for security
2. For a video diffusion model, what does temporal modeling add to the denoising process?
   • It guarantees every object remains physically correct
   • It lets the model represent dependencies among frames over time
   • It removes the image decoder from the pipeline
   • It turns each video into one text token
3. Why does the cited Stable Video Diffusion model use a latent representation?
   • To avoid using any learned decoder
   • To store all frames as discrete text tokens
   • To perform video diffusion on compressed latents before decoding frames
   • To guarantee there will be no flicker
4. How did the Stable Video Diffusion paper adapt an image diffusion model for video?
   • It replaced every image layer with a speech recognizer
   • It trained only a language model on video captions
   • It removed the image model and used optical flow alone
   • It inserted temporal layers into a pretrained image model and fine-tuned them on video data
5. What technique helps a video diffusion model follow a text prompt strongly?
   • Classifier-free guidance
   • Random shuffling of frames
   • Lowering the resolution
   • Removing the prompt entirely
6. Why are techniques like distillation and faster solvers being applied to video diffusion?
   • To make the model forget data
   • Because the many-step denoising sampling is slow and they reduce the number of steps
   • To add more noise permanently
   • To disable the text prompt
7. In the Stable Video Diffusion paper, what distinguishes the SVD-XT model from the 14-frame SVD model?
   • It replaces diffusion with autoregressive text generation
   • It removes temporal layers and generates only a still image
   • SVD-XT is fine-tuned to generate 25 frames instead of 14
   • It uses audio tokens as the only conditioning input
8. What is one tradeoff when using fewer denoising steps through distillation or a faster solver?
   • Zero compute is required
   • The video no longer needs conditioning
   • Every result becomes temporally perfect
   • Faster generation may change fidelity or motion quality and must be evaluated