This one started with a single question. How much can you strip away before the illusion breaks.

The goal was realism — not spectacle. Stable lighting, consistent character identity, and natural skin texture held across shots. Early versions tried to do more. Movement, paper planes, surreal transitions. The more complexity entered the frame, the more the illusion cracked.

The direction shifted toward reduction. Removing elements instead of adding them revealed what actually held the piece together. Stillness. Observation. Micro-expression. The final version relied on small shifts in gaze, posture, and breath rather than explicit action. Subtlety turned out to be the harder technical problem.

Firefly performed best when treated like a camera operator rather than a generator. Strong in lighting and composition, unreliable with complex motion or inanimate object behavior. Kling handled object physics more cleanly. Nano Banana produced the most consistent human results — facial identity and skin detail held better here than anywhere else in the stack.

The consistent pattern: the more layered the prompt, the less stable the output. One primary action per shot. Physical cues over emotional language. The system favors precision over complexity. Realism comes from constraint.