Hey team, are standard face recognition APIs actually holding up for you in production, or do they completely fall apart the second lighting changes? Looking for real experiences.