A four-second face replacement: her performance, someone else's head
Licensed footage of a rhythmic gymnast, and one studio portrait of a different person. The performance, the camera and the timing all stay hers; the head becomes theirs, hair included. The interesting part is not the swap, it is that a ribbon crosses her face twice in four seconds and her own forearm crosses it once.
Stage 1 - conform the plate to what the model accepts
Almost all of the work happens before anything is generated. The window is chosen for its occlusions rather than against them, the head is checked for movement by averaging every frame into one image, and the plate is cropped and relabelled to satisfy three separate hard limits on what the model will take.
Stage 2 - replace the head, hair included
A Seedance 2.0 node takes the conformed plate as the driving video and one cropped portrait as the identity. Because the new hair is short where the old hair was long, the model also has to invent the wooden wall and the white top that the long hair had been covering, and hold them steady for the whole move.
Step by step
- 01
Choose the window, and choose it for the occlusions
The source is a 6.6 second clip at 25 frames per second. Sampling every eighth frame shows the structure that matters: the ribbon and her hand cross her face repeatedly at the start, leave it clear through the middle of the clip, cross her mouth again, clear once more, then cross her nose at the end. The last frame is pure black and has to be avoided. The window picked opens on the clear run so the identity settles, then walks into two separate occlusion events. A face replacement is easy on the clear frames and only interesting on the others.
Every eighth frame of the source. The clear run is the middle row. - 02
Check whether the head moves at all
Averaging all 96 frames of the chosen window into a single image answers a question worth asking before reaching for a tracker. The head comes back sharp, which means it barely moves, while the ribbon smears into a soft band running straight down the middle of her face. So no head track is needed here, and the same picture says exactly where the occlusion is going to land.
96 frames averaged into one. Sharp means still. The vertical smear is the ribbon. - 03
Conform the plate to what the model actually accepts
Three constraints, none of them optional. The driving video is capped at 2,086,876 pixels per frame, and the untouched 2048x1080 source is 2,211,840, so it is rejected outright; a centre crop to 1920x1080 is 2,073,600 and fits. That crop is worth making for a second reason: 16:9 is the only aspect this model has recorded drift measurements for, so landing on it buys a known correction instead of an unknown one. Third, the model works at 24 frames per second and the source is 25, so 96 consecutive frames are relabelled rather than resampled. Every original frame survives, nothing judders, and the drive keeps a one to one mapping with whatever comes back.
- 04
The reference is one photograph
A single licensed studio portrait, 5792 by 8688, cropped down to the head. Almost everything about it disagrees with the plate: a hard studio key against soft warm bounce off pale wood, a teal backdrop spilling onto the skin, heavy blue eyeshadow, and a short auburn crop against long straight blonde hair. That last disagreement is the one that decided the whole approach.
The identity source, cropped from a licensed studio portrait. - 05
What we tried first, and dropped
The first plan kept the plate's hair and changed only the face, which meant the reference had to stop arguing with the plate. A structured JSON recipe restyled it: long blonde hair with darker roots, the makeup toned down, the studio key swapped for warm frontal bounce, four views on one sheet. It did every one of those things correctly and still lost the person, narrowing the jaw and softening the bone structure into a lookalike rather than a likeness. Two further misses were measurable: a background pinned to an exact grey came back about a tenth darker than specified, and head turns asked for at thirty five degrees came back far shallower. The sheet was abandoned when the brief changed to replacing the whole head, which deleted the problem it existed to solve. It is kept here because a reference that agrees with your plate is usually the right instinct, and because what it got wrong is more useful than what it got right.
The prompt for this stepShow promptHide
{ "task": "identity_preserving_restyle_to_reference_sheet", "reference_images": { "image_1": "portrait photograph of the subject. This is the identity source." }, "output": { "kind": "character reference sheet", "aspect_ratio": "1:1", "layout": "2x2 grid of four equal square quadrants, 8 px white gutter between quadrants, no labels, no text, no borders" }, "views": [ { "quadrant": "top_left", "head_yaw_deg": 0, "head_pitch_deg": 0, "framing": "head and shoulders", "gaze": "into lens" }, { "quadrant": "top_right", "head_yaw_deg": 35, "head_pitch_deg": 0, "framing": "head and shoulders", "gaze": "into lens" }, { "quadrant": "bottom_left", "head_yaw_deg": -35, "head_pitch_deg": 0, "framing": "head and shoulders", "gaze": "into lens" }, { "quadrant": "bottom_right", "head_yaw_deg": 0, "head_pitch_deg": 12, "framing": "head and shoulders", "gaze": "into lens" } ], "identity": { "source": "image_1, exactly", "instruction": "This must read as the same specific person as image_1, not a lookalike.", "lock_exactly": [ "face shape and proportions", "cheekbone height and width", "jaw width and chin shape", "eye shape, eye spacing, eyelid crease", "iris colour, blue-grey as in image_1, do not change", "nose bridge width, nose length, nostril shape", "lip shape, lip thickness, mouth width", "eyebrow shape and thickness", "skin tone and undertone", "apparent age" ], "consistency": "identical in all four quadrants. The same person, the same age, the same features." }, "change_from_reference": { "hair": { "length": "long, ending 10 cm below the collarbone", "style": "straight, centre parting, one section falling in front of each shoulder", "colour": "blonde, with darker brown roots for the first 4 cm from the parting", "forbid": ["short hair", "cropped hair", "pixie cut", "any hair above the jaw", "styled upward", "volume at the crown", "curls", "waves", "updo", "ponytail", "hair tucked behind the ears"] }, "makeup": { "eyeshadow": "neutral warm beige, matte, minimal", "liner": "thin, soft brown", "lip": "muted rose, matte", "forbid": ["blue eyeshadow", "smoky eye", "graphic liner", "red lip", "glitter", "heavy contouring"] }, "wardrobe": { "garment": "plain white high-necked short-sleeved top", "forbid": ["jewellery", "necklace", "choker", "earrings", "shoulder straps", "patterns", "logos"] }, "background": { "colour_hex": "#B3B3B3", "colour_rgb": [179, 179, 179], "description": "flat, even mid grey, identical in all four quadrants", "forbid": ["gradient", "vignette", "colour cast", "teal", "blue", "texture", "props", "cast shadow on the background"] } }, "lighting": { "key": "large soft frontal source, slightly above eye level", "colour_temperature_k": 4200, "fill": "warm bounce from below, as if reflected off pale wood", "contrast_ratio": "2:1", "forbid": ["hard shadows", "hard shadow edges", "coloured rim light", "backlight", "split lighting", "teal cast on skin", "blue cast on skin", "specular hotspot on the forehead"], "consistency": "identical lighting direction, intensity and colour in all four quadrants" }, "render": { "medium": "photoreal photographic portrait, 85 mm lens look", "focus": "sharp on the eyes in every quadrant", "skin": "natural texture with visible pores and fine lines", "forbid": ["retouching", "beauty smoothing", "airbrushing", "skin blur", "plastic skin", "stylisation", "illustration", "painting", "3d render", "added film grain", "over-sharpening", "text", "watermarks"] } }Correct on hair, light and wardrobe. Not the same face. - 06
Replace the whole head, hair and all
The plate carries motion, camera and timing. One reference image carries the identity. The prompt spends a single paragraph on what changes and all the rest on what must not: the ribbon's exact path through every frame it crosses her, her forearm staying in front of her face where it already is, the tattoo, the chain, the earrings, the panel spacing on the wall behind her, and the grain. Because the hair is now short where it used to be long, the model has the extra job of inventing the wall and the white top the long hair was covering.
The prompt for this stepShow promptHide
Replace this woman's entire head with the head of the person in the reference photograph. Her face and her hair both become theirs. Face: the bone structure, eye shape, eye spacing, blue-grey iris colour, nose, mouth, lip shape, jawline, chin, brow, and skin tone of the person in the reference photograph. This must read as that specific person, not a lookalike. Hair: short dark auburn hair, cut close at the sides and back, slightly longer and swept across the top, exactly as in the reference photograph. Her long blonde hair is gone. Because her hair is now short, reveal what the long hair was covering. The vertical wooden panelling continues uninterrupted behind her head and shoulders, and her white top continues uninterrupted across her chest and shoulders, both exactly as they would look if her hair had always been short. Her neck and the tops of her shoulders are now visible. Relight the new head to match this shot: soft warm frontal light with warm bounce coming up off the pale wooden wall. Remove the hard studio lighting and the teal colour cast of the reference photograph. Her new skin must sit in the same light as her neck, chest, and arms in this video. Everything else in this shot is unchanged, frame for frame. Camera: the same framing, the same lens, the same camera movement. Do not reframe, zoom, pan, crop, or stabilise. The edges of frame stay where they are. Timing: the same speed and the same motion throughout. Nothing slows down, speeds up, or holds. Head motion: her head position, rotation, tilt, and every small movement follow this video exactly. Her eyeline, her blinks, and where she is looking are unchanged. Performance: her expression is unchanged. The set of her mouth, the lift of her brows, and every micro-expression stay exactly as performed, now on the new face. Body: her neck, chest, shoulders, arms, hands, fingernails, and the tattoo on her forearm are unchanged. Wardrobe and jewellery: the white high-necked short-sleeved crop top is unchanged in shape, fit, and shading. The thin gold chain necklace and the gold hoop earrings stay exactly where they are. The ribbon: the pale cream gymnastics ribbon keeps its exact path, curl, width, and motion blur in every frame, including every frame where it passes in front of her. Do not remove it, redraw it, thin it, or change where it crosses her. Her forearm: where her forearm and hand pass in front of her body and in front of her face, they stay exactly as in this video, still in front, same pose, same position, same overlap. Background: the warm vertical wooden panelling is unchanged. Same panel spacing, same horizontal joint line across the wall, same soft defocus. Grade: the same contrast, the same warm colour, the same shallow depth of field, and the same grain. Photoreal live-action footage. No stylisation, no beauty smoothing, no skin retouching, no relighting of the body, no added sharpening.
Head, hair, and the reveal behind where the hair used to be, in one pass. - 07
What it got right, and what it cost
Right: the head is genuinely replaced rather than nudged, the identity reads closer to the photograph than the character sheet ever managed, the wall and the top behind the old hair are invented convincingly and stay put, the real ribbon survives and still crosses in front of her, the forearm stays in front, and the new head is lit by the room it is now in rather than by the studio it came from. What it cost: the return is 864 by 496, which is not 16:9 even though 16:9 was asked for, so it does not sit on the plate's geometry without being scaled back. It also hands back one frame more than it was given. Neither of those is worth complaining about; both are worth measuring and undoing before anything is composited on top.
How it was wired
Wired as: the conformed 1920x1080 plate to the Reference video input, the cropped portrait to Reference image 1, the prose prompt on a wired Prompt node, 4 seconds, 16:9. Nothing is composited afterwards. The result panel above is the model's own output, scaled back to the plate's 1920x1080 because the return does not come back at the aspect it was asked for.