Gids-gekoppelde quiz · Middelmatig Niveau
CLIP-score en quiz over menselijke voorkeursstatistieken
Test hoe de CLIP-score de uitlijning van prompts meet, hoe voorkeursmodellen zoals PickScore en ImageReward worden getraind en waar elke statistiek faalt.
Vraag 1 van 8
Waarom is FID alleen onvoldoende voor het evalueren van tekst-naar-beeldmodellen?
Bekijk elke vraag8 vragen
Antwoorden en uitleg verschijnen terwijl je de quiz doet.
- Waarom is FID alleen onvoldoende voor het evalueren van tekst-naar-beeldmodellen?
- It requires human annotators for every image
- It never looks at the prompt, so it cannot measure text alignment
- Het werkt alleen op zwart-witafbeeldingen
- It measures CLIP similarity instead of realism
- Hoe wordt een basis-CLIP-score berekend?
- By counting pixels that match the prompt's colors
- By training a discriminator on the generated images
- By averaging Inception features across a dataset
- By taking the cosine similarity between CLIP's image and text embeddings
- Which dataset was PickScore trained on?
- Pick-a-Pic, a set of choices made by users of a web app
- COCO captions only
- ImageNet class labels
- A collection of FID statistics
- Welke zwakte is typerend voor CLIP-gebaseerde scores?
- It cannot process color images
- It always penalizes photorealism
- It often misses counts, spatial relations and which attribute belongs to which object
- It requires prompts shorter than five words
- Waarom kunnen CLIP-scores van verschillende backbones niet rechtstreeks worden vergeleken?
- All backbones produce identical scores, so there is nothing to compare
- Different encoders such as ViT-B/32 and ViT-L/14 have different embedding spaces and score ranges
- CLIP score is only defined for one backbone
- Backbones only affect FID, not CLIP score
- What risk arises when a model is optimized directly against a preference score?
- The score becomes impossible to compute
- The model loses the ability to read prompts
- It may learn to please the scorer, for example with oversaturated images, rather than people
- FID automatically improves
- What typically happens to the part of a prompt beyond CLIP's 77-token limit?
- It is scored separately and averaged in
- The encoder expands its context window automatically
- It is summarized into fewer tokens first
- It is truncated, so it never affects the score
- Which kind of evaluation checks whether each object and attribute in a prompt actually appears?
- Question-answering or detector-based benchmarks such as TIFA and GenEval
- FID
- Raw CLIP cosine similarity
- Raising the guidance scale