A new room and outfit for a talking head: only Seedance 2.5 kept the person

The footage is the most common kind a creator has: a 48-second vertical selfie, phone held at arm's length, a plain room behind, a lilac T-shirt, natural gestures and speech in Russian. The goal was to put the same speaker in a warm evening living room, wearing a tailored blue blazer, without touching what makes the clip hers: the face, the hair, every gesture, the mouth shapes and the voice.

That rules out a lot. The viewer is looking straight at a face that talks, so a few frames of a mouth that does not match the words, or a face smoothed into someone slightly different, are enough to lose them. The clips below have no sound on purpose: the speech is in Russian, and what matters here is what you can see.

The process that worked

1. Pick a short test piece first. We cut five seconds (35.4–40.4 s) from the phone file, converted from 4K at 120 fps to 1080×1920 at 30 fps, with the sound kept. Test on five seconds before paying for forty-eight.
2. Prepare one reference still of the target look. A single frame of the same person already in the new outfit and the new room. One image can carry both the wardrobe and the location, as long as the prompt says it defines nothing else.
3. Run Seedance 2.5 in Video Edit mode. The source clip is the video reference, the still is the image reference, and the prompt is written as an edit order: what changes, what is the master, and what must not move. Generated sound is switched off.
4. Check the length. Our 5.00-second source came back as 4.71 seconds, its motion played about 6% fast. Stretched back to 5.00 s, the gestures lined up with the source within two frames.
5. Put the original voice back. The speaker's own recording goes on top of the retimed picture, so the voice is never a reconstruction.
6. Check the face at full size, then the lips with sound. Crop the face at 1:1 against the source before calling it done. Overlays, captions and inserts are added in the edit afterwards, never sent through the model.

Inputs you need

• The source clip, at least 4 seconds and at most 30 seconds per run (split longer takes).
• A reference still of the target outfit and room, ideally with the same person in it.
• The original audio, kept aside to be laid back on the result.
• The spoken language, so the mouth shapes follow the right phonetics.
• A plain list of what must survive (face, hair, hands, timing) and what must go (the old room, the old clothes, anything printed on the wall).

Route 1: restyle the real footage

This route keeps the recorded performance and redraws only what should change. We ran it on two Higgsfield models with the same five-second clip and the same reference still.

Source

Phone recording

The input: a plain room, a lilac T-shirt, natural gestures, hand-held selfie movement.

Reference

Target room and outfit

The reference both takes received: it defines the outfit and room, nothing else.

Take A · selected

Seedance 2.5 Video Edit

Room and outfit both changed; the face keeps its lines and skin texture, and the blinks, mouth shapes and hand gesture land where the source has them. We retimed the output before adding the original voice.

View prompt

Take B

Higgsfield Genjutsu

Motion transfer copies the performance closely and adds the blazer, but the old room stays, and the face comes out smoother and less like the source. Genjutsu takes no prompt.

Seedance 2.5 is the model to use when a real person stays in frame. It was the only one that changed the room and the clothes while keeping a face you would recognise at full size. Genjutsu did half the job and softened the face. We also upscaled the Genjutsu clip from 720p to 1080p with ByteDance Video Upscale: the upscaler invents lashes and texture, and on a face that had already drifted it made the result look worse, not better. Upscaling does not fix a weaker model.

The mode matters as much as the model. Our first two Seedance 2.5 attempts used the reference (omni) mode with the clip attached as a video reference. Both failed after about 45 seconds and were refunded, with no reason given through the API. The Higgsfield web page gave the answer: Your prompt looks like a video editing request. Switch to the Video Edit mode so the result follows your input video. In Video Edit mode the same job went through on the first try. Video Edit bills by the length of the source clip: 38 credits for five seconds at 720p.

Route 2: cut the person out and composite

The cheaper route never redraws the person. A matting model finds which pixels belong to the speaker; the original pixels are then laid over a new background with ffmpeg. We compared three providers on a seven-second piece of the same recording, with a city view from a window as the new background. This piece carries three small inserted video overlays, shown as they appeared in the input.

The weak point of every cut-out is the same: fine hair against a bright wall. Each result keeps some of the old room between the strands, which reads as a halo against a new sky. Inserted overlays are the second trap: a matting model only knows the person and everything else, so keep overlays out of the clip you send and add them back in the edit.

Route 3: regenerate the speaker from one photo

The third route throws the footage away and animates a single photo with the original voice. We gave two avatar models the same still of the speaker at a lakeside marina and the first 5.55 seconds of her recorded speech, ending on a natural pause.

Source photo

Marina portrait

The single photo both avatar models animated, with the speaker’s own voice.

Avatar 1

OmniHuman 1.5

Lively head and mouth movement, but the skin is smoothed and she looks slightly younger. $0.89 on fal, 3 minutes.

View prompt

Avatar 2

Kling AI Avatar v2 Pro

Closer to the photo, with freckles and texture kept and a believable hand-held sway. $0.83 on fal, 7 minutes.

Both keep the real voice, because they only animate the face to it. What they cannot keep is the real performance: every gesture and head turn is invented. That suits a new piece to camera, but it is a different video, not the original one in a new room.

The template

This is the full restyle template in its most general form: it can replace the room, the wardrobe and the hair, keep chosen props from the source, and re-cut one take into several angles. Fill the bracketed slots and delete what your job does not need. On Higgsfield, submit it in Seedance 2.5 Video Edit mode.

• Keeping the speaker's own hair: make @Look wardrobe only.
• Nothing to keep from the old room: delete the props line.
• One continuous take, such as a selfie: use a single stage, and delete the lines about new angles and shot sizes.
• Phone footage: replace the cinema-camera style line with "photoreal phone video matching the source", or the result will not cut with the rest of the recording.
• The original voice: switch generated sound off and lay the recording back afterwards.

[Edit Goal]
Edit @Video 1. For the entire video, replace the room behind the woman and replace her lilac T-shirt with a blazer outfit.

[Source Video Role]
@Video 1 is the sole editing master. It defines the woman: her face, skin, freckles, eye colour, hair, expression, every blink and every mouth shape of her Russian speech, her hand gestures, her position and size in frame, the handheld selfie camera movement, the audio, and the timing of everything.

[Target Material Role]
@Image 1 defines only two things. Her clothing: a tailored dusty-blue blazer with wide peak lapels over a matching V-neck waistcoat with buttons. The room: a cosy warm living room in the evening, a large framed landscape painting in gold, ochre and soft blue on the left, a table lamp with a cream shade glowing on a wooden sideboard, built-in shelves on the right lit by warm strip lights with dark ceramic vases and books. Do not use the woman's face, hairstyle or pose from @Image 1.

[Edit Scope]
Modify only the background outside her silhouette, and her clothing below the neck. Allow the warm lamp light of the new room to fall naturally on her face, hair and blazer. Do not modify her identity, facial features, freckles, eye colour, hairstyle, expression, mouth movement, hands, position, size or motion.

[Timeline Inheritance]
The blazer moves exactly as the T-shirt did, following every shoulder and arm movement, and her hands pass in front of it at the same moments. Keep all her actions, gestures and the camera movement from @Video 1 unchanged, frame by frame. Exactly one person. No on-screen text or logos.

[Content to Preserve]
Keep her face, hair, expression, lip movement, gestures, camera framing and sway, timing and the original audio from @Video 1.

A woman talks directly to the camera in a handheld selfie video on a sunny lakeside marina. Natural, relaxed delivery: small head nods and tilts, lively eyebrows, natural blinks, subtle shoulder movement, slight handheld camera sway as if she holds the phone at arm's length. Her hair moves lightly in the breeze; boats rock gently on the water behind her. Her face, hair and striped shirt stay exactly as in the photo. Realistic skin, natural daylight.

Inputs: the marina photo as the image, the speaker's own recording (0–5.55 s) as the audio. OmniHuman 1.5 at 720p.

@Source is <SOURCE>, the master motion track. This is a restyle of <SOURCE>, not a new performance: <HERO>'s face, body, pose, every movement of head, hands and fingers, every gesture, every object touched, speech timing and mouth shapes are copied from <SOURCE> frame by frame, at exactly the same seconds, same speed, same direction, same hand. Ignore from <SOURCE>: [FILL IN: what to remove from the source video, such as the old room, furniture, clothing, hairstyle].
<SOURCE> also defines <PROPS>, identical in every shot: [FILL IN: objects from the source to keep, each one separately: what it is, colour, material, where it stands. IF THERE IS NOTHING TO KEEP, DELETE THIS WHOLE LINE]. Exactly one of each object.
@Location is <LOCATION>: [FILL IN: the new location: furniture, where the hero is, lamps and windows, walls, decor]. Carry this geometry identically. Ignore its camera angle.
@Look is <LOOK>, wardrobe and hair only: [FILL IN: clothing from top to bottom, accessories, shoes; hairstyle: length, shape, styling]. Not the face. Ignore the background and any text on the reference.

[Blocking]
<HERO> takes the same position as in <SOURCE>, now [FILL IN: where the hero is in the new location, for example sitting in an armchair or standing by the window], with [FILL IN: what is behind the hero] behind. Scale and distances stay as in <SOURCE>.

[Stages]
Stage 1 ([FILL IN: timecode, for example 0–4] s) | Camera: [FILL IN: shot size, angle, static or moving camera]. <HERO> [FILL IN: what the hero does] exactly as in <SOURCE>, talking with every word.
Stage 2 ([FILL IN: timecode] s) | Camera: [FILL IN: shot size, angle, static or moving camera]. <HERO> [FILL IN: what the hero does] exactly as in <SOURCE>, talking with every word.
Stage 3 ([FILL IN: timecode] s) | Camera: [FILL IN: shot size, angle, static or moving camera]. <HERO> [FILL IN: what the hero does] exactly as in <SOURCE>, talking with every word.

[Maintain consistency]
Every movement matches <SOURCE> at the same timecode; nothing is added, skipped, slowed or sped up. In new angles, the same movements are seen from the new angle. <HERO> talks exactly as in <SOURCE> in every shot, the mouth never still while <SOURCE> has speech. Exactly one person. The face matches <SOURCE> from every angle, [FILL IN: any change to the face, for example fully clean-shaven, OR DELETE], with visible pores and dry matte skin, never plastic. Wardrobe and hair match <LOOK> in every shot. Every shot differs in size or angle from the one before. Static stages have no pan, no zoom and no drift. No on-screen text, subtitles or logos.

[Lighting]
[FILL IN: time of day, for example Evening, no daylight]. Key: [FILL IN: the main light source: what it is, where it is in frame, its colour]. Fill: [FILL IN: the second light source: what, where, colour]. Light and shadows on <HERO> come from the light sources of <LOCATION>, with soft contact shadows. <HERO> belongs to the room, never pasted over it.

[Style]
Photoreal live action, real cinema camera, shallow focus, natural motion blur, [FILL IN: colour grade, for example a warm cosy grade], fine film grain. No CGI look, no sharpening, no HDR.

[Audio]
Dialogue language: [FILL IN: the spoken language]. The soundtrack is the original audio of <SOURCE>, unchanged: same voice, words, intonation and timing, with mouth shapes matching [FILL IN: the spoken language] phonetics. No music, no added sounds.