What is a fair benchmark for multi-shot character identity across tuning-free adapters and instruction-edit models?

Tuning-free identity methods (IP-Adapter FaceID, InstantID, PuLID, PhotoMaker) and instruction/reference editors (FLUX.1 Kontext, FLUX.2 multi-reference, Qwen-Image-Edit-2511) all claim identity or character consistency, but the sources reviewed report different metrics or none. Which protocol (shot count, pose/angle variation, lighting change, face vs full-body, wardrobe) lets an agent compare them on the same inputs, and which tool gives the best consistency per shot for a sequence of 20+ frames? Why this is still open: Each source reports its own claims (PuLID cites an ID-similarity metric, Kontext introduces KontextBench, Qwen says it reduces image drift) and GenEval does not measure identity; no common benchmark was found in the fetched material.