Image-to-Video Generation Quiz

Answers and explanations appear as you take the quiz.

1. In many latent video diffusion models, how is the input image given to the denoiser?
   • As a text caption generated from the image
   • Its latent is concatenated channel-wise with the noisy frame latents at every step
   • It is used only after generation to color-correct frames
   • It replaces the last frame of the clip
2. What does Stable Video Diffusion's motion bucket value represent?
   • The number of frames
   • The output file size
   • The amount of motion, learned from optical flow in training clips
   • The camera's focal length
3. What happens when you raise the noise augmentation on the conditioning image?
   • More motion and freedom, but less fidelity to the input
   • The video becomes perfectly still
   • Resolution doubles
   • The first frame matches the input exactly
4. Why does the area revealed behind a moving object often look invented?
   • The model deletes background pixels on purpose
   • Compression removes background detail
   • The input image is cropped
   • The model has no information about occluded regions and predicts plausible pixels
5. What is the main drawback of extending a clip by feeding its last frame back in as a new input?
   • It is not allowed by any model
   • Errors accumulate, so quality degrades with each extension
   • It always reverses the motion
   • It removes the audio
6. What is a good prompting practice for image-to-video?
   • Restate every object visible in the image
   • Leave the prompt empty always
   • Describe the motion and the camera rather than restating the scene
   • Use only a single noun
7. Which description matches the artifact called texture swimming?
   • Fine patterns appear to slide or shimmer across surfaces
   • The whole video turns black and white
   • Audio drifts out of sync
   • The clip plays backward
8. How does Stable Video Diffusion's reference pipeline apply classifier-free guidance across frames?
   • Highest on the first frame, then off
   • Randomly per frame
   • Identical on all frames
   • Ramped linearly, lower near the conditioning frame and higher for later frames