Quiz collegato alla guida · Medio Livello

Punteggio CLIP e quiz sulle metriche delle preferenze umane

Verifica come il punteggio CLIP misura l'allineamento dei prompt, come vengono addestrati i modelli di preferenza come PickScore e ImageReward e dove ciascuna metrica fallisce.

Domanda 1 di 8

Perché il FID da solo non è sufficiente per valutare i modelli text-to-image?

Anteprima di ogni domanda8 domande

Le risposte e le spiegazioni vengono visualizzate mentre rispondi al quiz.

  1. Perché il FID da solo non è sufficiente per valutare i modelli text-to-image?
    • It requires human annotators for every image
    • It never looks at the prompt, so it cannot measure text alignment
    • Funziona solo su immagini in bianco e nero
    • It measures CLIP similarity instead of realism
  2. Come viene calcolato il punteggio CLIP di base?
    • 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
  3. Su quale set di dati è stato addestrato PickScore?
    • 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
  4. Quale debolezza è tipica del punteggio basato sul CLIP?
    • 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
  5. Perché i punteggi CLIP di dorsali diverse non possono essere confrontati direttamente?
    • 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
  6. 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
  7. 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
  8. 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