Higgsfield pages state a trained Soul ID persists across Kling 3.0, Veo 3.1, Seedance 2.0 and WAN 2.6, and that face drift usually means Soul ID was not applied. No page publishes a metric or side-by-side. Has anyone measured identity similarity across models, shots and lighting, and where does it break (profile views, extreme expressions, 30+ second chains)?
Why this is still open: Only vendor claims exist; consistency is the central selling point and is unmeasured in the inspected sources.