Higgsfield Genjutsu has two modes, and they do very different jobs. If your background drifts or the light looks wrong, you're almost certainly in the wrong one.
Genjutsu rebuilds the person frame by frame, so the cleaner your inputs, the less it has to guess. These make the biggest difference.
I've attached two files: a photo of a character (Image 1) and my video (the source video). Use Higgsfield Genjutsu in Object Swap mode to replace me with the character. Do not use Motion Transfer, because it rebuilds the scene and changes the background and lighting.
Before generating, check both files and tell me in three short bullet points if anything will hurt the result: the video should show one person in one continuous shot with a steady camera and be under 30 seconds; Image 1 should show one person with a clear, front-facing face, framed the same way as the video, with hands visible if my hands are visible; and our builds should be roughly similar. If something is off, ask me whether to continue.
Generate a 480p test first, sending Genjutsu this prompt exactly as written:
Replace only the person in the source video with the character from Image 1. Everything else in the frame stays exactly as it is in the source video.
Identity: the character's face, hair, skin tone and body build come 100% from Image 1, in every frame. Outfit: [the outfit from Image 1 / keep the original outfit from the source video]. The same face, hair and outfit hold from the first frame to the last, including in profile, mid-blink and mid-word.
Performance: the character copies the original performance frame for frame. Every head turn, eyebrow raise, blink, glance, hand gesture, shoulder movement and pause happens at the same moment, with the same size and speed. Mouth shapes follow the original speech exactly, so the lip sync lines up with the original audio. Hands keep five fingers and follow the original hand positions and anything they touch.
Scene: the background is identical to the source video in every frame: the same walls, furniture, objects, screens and depth of field, all in the same positions and colours. The camera framing, crop, focal length and any camera movement stay exactly as in the source.
Lighting: the character is lit by the same light as the original person: the same key light direction, the same softness, the same shadow side of the face, the same colour temperature and exposure. Shadows the character casts on nearby surfaces match the original person's shadows. Brightness and white balance stay constant across every frame.
Edges and texture: the character fully covers the original person's silhouette, with clean, natural edges on hair, ears, shoulders and fingers. Skin keeps real texture, and the character matches the video's sharpness, grain and compression, so they look filmed in the room rather than pasted on.
Locks: one continuous shot at the same length and frame rate as the source. The background, lighting and camera stay identical to the source in every frame. The character's identity from Image 1 stays identical in every frame.
Avoid: any trace of the original person (face, hair, clothing, outline or skin showing through), halos or shimmer around the edges, the face or body morphing between frames, flickering light or colour shifts, the background warping, sliding or melting, extra or missing fingers, smeared hands, lip movement out of time with the audio, plastic or airbrushed skin, added objects, text or watermarks, and any new zooms, camera moves or cuts.
After the test, check it against this list and tell me what failed: identity holds in every frame, background unchanged, lighting and colour constant, gestures and lip sync in time, clean edges and hands. For each failure, add one line to the Genjutsu prompt naming exactly what went wrong and how it should look instead, then run the test again. Once the test passes, run the same prompt at 1080p.This is the version I use on a mic-in-hand talking-head clip when the background must not move at all: the room, the edges around the person, the lighting. It treats the swap as a person inside a fixed plate. The video is the master for everything, the photo only supplies the identity. Change only the bracketed line for each person you swap in.
SCENE CONTEXT
Replace only the woman in @video1 with the person from @image1. The background, the camera, the framing, the lighting, the timing, the motion and the audio of @video1 stay exactly as they are. Treat this as a person swap inside a fixed plate: every pixel outside the person's outline is copied from @video1 unchanged, and the new person is composited into that outline with the same edges, the same light and the same motion.
ACTIVE REFERENCES
@video1 = the master for everything except identity. Copy its background, camera, framing, motion, lighting, timing and audio frame for frame.
@image1 = the new person's identity only: face structure, skin tone, eye shape and colour, eyebrows, nose, lips, hair colour, length and texture, body build. [ONE LINE: age, hair, skin tone, build]. 100 percent matches the reference. Nothing of the original woman's face, hair or skin carries over.
BACKGROUND PLATE (locked, zero change)
Vertical 9:16 phone frame. Warm cream ceiling filling the top third with a round gold ceiling-light fitting above and behind the head. Cream walls. On the left, an arched doorway showing a warm lit lamp on a side table and beige curtains behind it. On the right, a dark wooden cabinet with a vase and a framed picture. A soft grey sofa edge in the lower left corner. Every one of these elements keeps its exact position, size, colour, brightness, sharpness and grain in every frame. Nothing is repainted, regenerated, cleaned up, blurred, sharpened, shifted or re-lit. The soft focus falloff of the room behind the subject is identical to @video1. Where the new person's silhouette is thinner than the original, the revealed background is filled with the true background from neighbouring frames of @video1, matching the same wall, arch and cabinet, never invented.
EDGES (the hardest part, treat with care)
The boundary between the person and the room is clean and natural in every frame: hair edges show individual strands and flyaways against the wall, with no halo, no glow, no dark outline, no smeared or ghosted rim, no double edge, no jagged cut-out, no soft blur band around the body. The shoulders, the neck, the ear with the earring and the edge of the black top meet the background exactly where they do in @video1. The hand and the white microphone keep their exact outline and their exact overlap with the chest. Shadows the body casts onto the top and the neck stay consistent with the room light. No flicker or shimmer along the outline between frames.
FIRST FRAME / BLOCKING
The new person occupies the same silhouette, head position, head tilt and shoulder line as the woman in @video1's first frame: head slightly right of centre, tilted a little to camera-left, eyes on the phone lens, mouth open mid-word. Left hand holding a small white wireless microphone with a white fluffy windshield and a cartoon sticker on the transmitter at chest height, fingers curled around it in the same grip.
FORMAT MODE
One continuous shot, the camera does not cut on its own. Same duration as @video1, same frame rate, same 9:16 vertical framing, no zoom, no reframing, no crop, no stabilisation.
CAMERA (locked from @video1)
Front phone camera at eye level, roughly 84 degree wide selfie-lens perspective with its slight wide-angle stretch. Same static handheld micro drift as @video1. Focus on the face. Exposure, white balance, depth of field and grain identical to @video1 in every frame.
ACTION (locked from @video1)
The new person performs the exact motion of @video1: every head turn, nod, shoulder shift and hand movement with the microphone at the same frame, the same speed and the same amplitude. Lip movement follows @video1's mouth shapes syllable for syllable so the original audio stays in sync. Blinks land where @video1 blinks. Eye line follows @video1's eye line.
PERFORMANCE
Same energy and beats as @video1 (talking to camera, warm, mid explanation, small smile on the same words), performed with the new person's face. Skin at close phone range: visible pores, natural sheen on the forehead and nose, real catch-lights in the eyes from the lamp and the front light, fine flyaway hairs at the hairline. Teeth, tongue and inner mouth are the new person's own.
HAIR AND BODY
The new hair moves with the head turns the way real hair does and settles naturally, never freezes, floats, or clips through the shoulders or the microphone. Neck, hands and every visible patch of skin match the reference face's tone, including the hand holding the microphone.
WARDROBE (locked from @video1)
Same black long-sleeve V-neck wrap top with the tie at the front, same white top underneath, same small earring in the same place, same microphone, same sticker, same windshield. Fabric folds move exactly as in @video1.
LIGHTING (locked from @video1)
Warm indoor practical light around 3200K: soft key from the front and slightly camera-left, the warm lamp glow through the arch on the left, gentle top light from the ceiling fitting. The new face is lit by exactly this light with the same shadow under the chin, the same highlight on the forehead and nose, the same soft falloff on the camera-right cheek and the same warm colour cast on the skin. No new light source, no ring light, no studio fill, no change to the room's brightness.
AUDIO
Original @video1 audio, untouched, in sync with the new mouth.
STYLE
Unedited phone video straight off the camera. Same resolution, same grain, same slight softness as @video1. No colour grade, no beautification, no smoothing, no sharpening, no upscaling.
POSITIVE LOCKS
Background, ceiling light, arch, lamp, curtains, cabinet, vase, picture and sofa are pixel-identical to @video1 in every frame. Camera position, framing, focal length, exposure and grain are identical to @video1 in every frame. The person's outline sits exactly where it does in @video1 with clean natural edges and no halo. Every movement matches @video1 at the same frame. Microphone, hand grip, sticker, windshield and wardrobe are identical to @video1. The face, hair, skin tone and body are @image1's from the first frame to the last: the same person in every frame, no drift, no morphing, no blend of the two identities, no frame where the original woman reappears. Lip sync stays locked to the original audio for the full clip.Same trick, but the "character" is you. You do not need a photoshoot: take a screenshot of yourself from the video, ask ChatGPT to glam it up while keeping everything else identical, then hand that photo to Genjutsu as the reference. Because the glam photo has the same framing, background and light as the video, the swap holds far better than a random photo would.
Take the attached screenshot of me from my video and create a glammed-up version of myself. Same person: same face structure, same eye shape, same nose, same lips, same skin tone, same body build, same head position, same tilt, same expression, same framing, same background and same lighting as the screenshot, so it still matches my video frame for frame.
Upgrade only the styling. Hair styled in [soft Hollywood waves / a sleek blowout / a polished updo], full glam makeup with defined brows, soft-glam eyes with lashes, a lifted contour and a satin lip in [colour]. Skin still looks like real skin at phone range, with pores and a natural sheen on the forehead and nose, never airbrushed. Outfit: [keep the same outfit / a dressier version of the same outfit in the same colours]. Keep my earring and the microphone in my hand exactly where they are.
Photoreal, front phone camera quality, the same warm indoor light from the same direction with the same shadow under the chin, no filter look, no plastic skin, no change to the room, no change to my face shape or features, no text, no watermark.Pricing from Higgsfield's Genjutsu announcement.
🎭 Higgsfield
📖 Genjutsu modes · announcement and pricing
📜 Terms of use