Read this when a slideshow uses recurring people, mirror selfies, narrator-owned POV, or an iPhone camera-roll finish.
An image reference can improve consistency, but it does not replace a complete prompt. Each request must independently resolve:
- character identity, stable facial markers, hair, body, hands, nails, jewelry, and retouching limits;
- exact wardrobe for that frame;
- device, case, lenses, camera app, lens, and disabled modes;
- room geometry, fixtures, soft goods, clutter limits, lighting, and capture artifacts;
- camera owner, optical path, position, angle, body visibility, and phone visibility;
- anatomy and quality rejection criteria.
Never use same woman, same room, as before, previous image, or similar prose as a substitute. scripts/build-prompt-package.py resolves named profile entries into every prompt and rejects stateful shorthand in frame-critical fields.
The narrator holds the phone and aims its rear camera into a real mirror. The phone, holding hand, and forearm are visible only as a reflection. The body faces the mirror and the eyes look toward the reflected phone screen. Reject missing or duplicated phones, external photographer geometry, front-camera framing disguised as a mirror selfie, impossible handedness, and reflections that do not match the room.
The image is exactly the narrator's phone-screen view. The phone and camera-holding hand occupy the camera position and cannot appear. Only physically possible anatomy may enter the frame: the free hand or forearm, lap, knees, legs, socks, shoes, or a small torso edge.
Keep a free forearm near the camera. It should enter from a nearby lower corner and end within the near half of the scene. A hand reaching from the near edge to a far-third object produces exaggerated limb length and must be rejected even when the rest of the frame looks convincingly iPhone-shot.
Use the smallest relevant reference set. State what each image controls—identity, product geometry, room, or finish—and what it must not copy. The stateless written profile remains mandatory. Approve the anchor before paying for dependent frames.
The same phone cannot photograph its own screen. Use deterministic screenshots, screen recordings, or a separately photographed verified device when product proof requires exact UI. Mark those slides generation.render_mode: deterministic_asset; the prompt builder emits an asset contract rather than asking Muse to invent proof.
At native resolution reject:
- extra, fused, or malformed fingers;
- elongated or rubber-like forearms;
- a hand larger than its plausible distance permits;
- detached limbs or impossible joints;
- duplicated phones or hands;
- camera-owner geometry that implies an unseen photographer;
- mirror handedness or optical paths that cannot exist.
Do not excuse anatomy failures as casual authenticity.