Evaluation¶

Compute quality metrics comparing real and generated image distributions.

Metric Types¶

Distribution-To-Distribution (DTD)¶

Measure overall similarity between the real and synthetic distributions.

Metric	Description
`fid`	Fréchet Inception Distance : lower is better
`inception_score`	Inception Score : higher is better

Point-To-Distribution (PTD)¶

Score each individual synthetic image against the real distribution.

Metric	Description
`mld`	Mahalanobis Distance : lower means more similar to real distribution

Usage¶

from ciagen import evaluate

scores = evaluate(
    real="data/real/train/images/",
    generated="data/generated/",
    metrics=["fid", "mld"],
    feature_extractor="vit",
)

Feature Extractors¶

Both DTD and PTD metrics operate on deep feature representations, not raw pixels.

Extractor	Model	Dimensions
`vit`	ViT-base (default)	768
`inception`	Inception v3	2048

CLI¶

ciagen evaluate \
    --real data/real/train/images/ \
    --generated data/generated/ \
    --metrics fid mld \
    --feature-extractor vit