Heru Chan: GPT Image 2 Generated Synthetic Person and Character Consistency

This is Heru Chan. GPT Image 2 generated Synthetic person. character consistency is the hard part of ai image gen. one good image is easy. the same face across 10 scenes is not. the model drifts every generation. it beautifies, smooths skin, quietly changes the face. i built a fictional character to solve this. built on gpt image, next test is rebuilding it local on my 4060. here's what actually held identity together: identity lock: a tight facial description. structure and features, not beauty language. beauty words make the face generic. negatives matter as much as positives. "no drift, no over-retouch, no exaggerated symmetry" cut inconsistency more than adding detail did. production clarity: flat lighting, neutral background, eye-level framing. cinematic light looks nicer and breaks consistency faster. scene restraint: the more complex the scene, the more the face drifts. lighting, angle, outfit all fight the face for control. too little detail, the face drifts. too much, the model gets confused. the balance is identity lock + production clarity + scene restraint. this is where prompt engineering becomes character engineering. full prompt in the comment.