Which AI video model keeps one character consistent across scenes?

Create an 11-second cinematic sequence with exactly three clean cuts and one recurring fictional character, Mara Vale. Preserve her identity across every shot: warm olive skin, angular oval face, hazel-green eyes, short diagonal scar at the outer left eyebrow, chin-length asymmetrical dark-auburn bob with a narrow silver streak at the left temple, slate-blue cropped utility jacket, burnt-orange scarf, charcoal technical trousers, black ankle boots, and teal triangular pin. Do not change her face, age, hair, wardrobe, body proportions, or accessories between shots. Shot 1, 0-3.5 seconds: dawn on a windswept mountain observation platform. Wide-to-medium view. Mara adjusts a compact field sensor, then looks toward a distant ridge. Slow controlled dolly inward. Shot 2, 3.5-7.5 seconds: hard cut to a dim greenhouse laboratory. Medium profile view. Mara studies a translucent signal map, traces one line with her finger, then glances toward camera. Gentle lateral camera move. Shot 3, 7.5-11 seconds: hard cut to a rocky field at sunset. Medium close-up. Mara lowers a handheld scanner, turns toward camera, and finishes the final line with calm certainty. Subtle push-in. Use IMAGE_REFERENCE as Mara's identity and wardrobe reference. Use AUDIO_REFERENCE as the exact voice, cadence, dialogue, and timing as the exact voice, cadence, dialogue, and timing. Lip-sync Mara naturally when her face is visible. Preserve the original spoken words. No additional dialogue, narrator, music, or captions. Add only restrained location ambience beneath the supplied voice. Natural cinematic lighting, realistic skin and fabric, grounded movement, no glamour retouching, no text or logos.

Only the reference tokens changed: Wan and H3 used Image 1 and Audio 1, while Seedance used @Image1 and @Audio1. The repeated phrase near the end was an input typo and was deliberately left unchanged across all three requests.