Available evaluators
Face similarity compares the largest detected face with a normalized centroid of your reference embeddings. Generic identity excess subtracts the exact matched base score. Detection misses are counted separately.
Image RMSE measures pixel change, not semantic quality. You can also import scalar measurements from CSV or provide a Python evaluator plugin for other concepts, representations, adapters, or models. Behavioral measurements alone do not identify a mechanism.
Local quick start
From a DeTangle checkout:
python -m pip install "./lab[face,images]" deltascope import-csv images.csv --out dataset.json --protocol protocol.json deltascope references dataset.json --folder refs deltascope serve dataset.json --face-model-dir models/buffalo_l
Model weights and reference photos are supplied locally. No model download is triggered by viewing or scoring. InsightFace code and pretrained weights have separate licensing terms; see the package notices before choosing weights for your use.
Export for GitHub Pages
deltascope export dataset.json --out site
Exports the selected renders, prompts, settings and scores to a static site. Scoring remains local.