---

title: evaluate() description: Python API reference for ciagen.evaluate() : compute FID, Inception Score, and Mahalanobis distance for generated images. keywords: api, evaluate, python, fid, inception score

---

evaluate()

Evaluate the quality of generated images against real images.

from ciagen import evaluate

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

Parameters

Parameter	Type	Default	Description
real	str \| Path	required	Directory containing real images
generated	str \| Path	required	Directory containing generated images
metrics	list[str] \| None	["fid", "mld"]	Metrics to compute: fid, inception_score, mld
feature_extractor	str	"vit"	Feature extractor: vit or inception
batch_size	int	32	Batch size for computation
limit_size_real	int	2000	Max real images to use
limit_size_syn	int	2000	Max synthetic images to use
image_formats	list[str] \| None	["png","jpg","jpeg"]	Supported formats
device	str \| None	auto	"cuda" or "cpu" (auto-detected if None)

Returns

dict with two optional keys depending on which metrics were computed:

dtd key (Distribution-To-Distribution)

{
    "fid": {
        "vit": 45.23
    },
    "inception_score": {
        "vit": 12.5
    }
}

ptd key (Point-To-Distribution)

{
    "mld": {
        "vit": {
            "/path/to/image_1.png": 2.34,
            "/path/to/image_2.png": 3.56,
        }
    }
}