What problem does it solve? After style transfer, inpainting, or background replacement, a subject's face or pose can drift away from the original. This Skill provides a deterministic pass/fail check that compares a candidate image's metadata against a reference so identity drift is caught before a workflow step is marked done. ## Core Features & Use Cases - Heuristic identity scoring: Combines face bounding-box IoU, subject token overlap, and goal-aware denoise checks into a 0-1 score with a 0.6 pass threshold. - Structured JSON contract: Reads {reference_meta, candidate_meta, constraints} on stdin and emits {pass, score, reasons[], evaluator} on stdout, making it easy to swap in a real face-embedding comparator later. - Fail-safe behavior: Returns pass: false when required structure is missing instead of guessing, so callers can escalate or trigger a repair loop. - Use Case: After running a style-transfer step on a portrait, pipe the reference and candidate metadata into the evaluator to confirm the face region and subject tokens still match before accepting the result. ## Quick Start Run the identity preservation evaluator on the reference and candidate metadata from my last style-transfer step and tell me whether the subject's identity was preserved.
