An AI-generated-but-consistent control to disentangle "detecting AI generation
artifacts" from "detecting claim-image inconsistency". For each TARA test image we
apply, via the FLUX.2-klein-9B reference pipeline (seed 42, 28 steps), a benign,
context-appropriate edit that does NOT break consistency with the claim — the same
generator and edit magnitude as the inconsistent set, the only difference being whether
the edit… See the full description on the dataset page:
https://huggingface.co/datasets/zengrh3/flux2-consistent-pilot.
