Gids-gekoppelde quiz · Middelmatig Niveau
AI-videogeneratoren evalueren met VBench Quiz
Test de 16 dimensies van VBench, hoe beoordelaars zoals DINO-functies en optische stroom video's scoren, en waarom menselijke arena's een aanvulling vormen op benchmarks.
Vraag 1 van 8
How many evaluation dimensions does the original VBench define?
Bekijk elke vraag8 vragen
Antwoorden en uitleg verschijnen terwijl je de quiz doet.
- How many evaluation dimensions does the original VBench define?
- 4
- 8
- 16
- 32
- Which dimension would reveal that a model produces nearly static videos?
- Dynamic degree
- Background consistency
- Aesthetic quality
- Color
- What does VBench use to measure subject consistency across frames?
- Pixel histograms
- FID statistics
- Audio embeddings
- DINO features
- Which two families group VBench's dimensions?
- Speed and cost
- Video quality and video-condition consistency
- Realism and diversity
- Training en gevolgtrekking
- Waarom kan het alleen lezen van VBench’s totaalscore misleidend zijn?
- The total is always zero for open models
- It measures only audio
- A timid, nearly frozen model can score high on consistency while barely moving
- It ignores every dimension except color
- How is motion smoothness estimated?
- By checking how well a frame interpolation model reconstructs dropped frames
- By counting frames per second
- By asking users to rate each clip
- By measuring file size
- What does VBench-2.0 focus on?
- Resolution only
- Deeper properties such as physics, commonsense and human anatomy
- Audio synchronization only
- Replacing all evaluators with FID
- Why do human video arenas still matter?
- They are cheaper than automatic metrics
- They are perfectly reproducible
- They measure only frame sharpness
- They capture overall preference, including qualities such as plausible motion that evaluators miss