← AI Terminology
CLIP Score
CLIP Score measures alignment between an image and a text caption using CLIP embeddings’ similarity — higher means better text–image match.
Common for text-to-image evaluation and filtering.
Common for text-to-image evaluation and filtering.
Why It Matters in AI
FID ignores prompts; CLIP Score checks whether the image matches the text. Used to evaluate T2I models, rank samples, and filter noisy image–text web data for training.
Key Points
| Aspect | Description |
|---|---|
| Use | T2I eval, rerank samples, data filtering |
| Limits | CLIP’s own biases and failure modes |
| Method | cosine(CLIP_image(x), CLIP_text(t)) |
| Related | CLIP, FID, human preference (PickScore, etc.) |
| Variants | RefCLIPScore with references |
| Training data | Keep high CLIP-score pairs from the web |
Simple Analogy
A bilingual critic scoring how well a poster matches its slogan — not whether the poster looks photographic alone.
Common Usage Examples
- Rank 8 diffusion samples by CLIP Score
- Filter LAION-style datasets
- Report CLIP Score beside FID
- Watch gaming via adversarial images
Summary
In short: CLIP Score rates text–image alignment via CLIP similarity — the go-to automatic metric for prompt fidelity in image generation.