Part of the Changeling family, all together in Changeling Studio: Imago · Imago Realtime · Tiresias (gender) · Chronos (age) · Natio (where in the world) · Cosmeta (makeup, hair, beards) · Echo (voices) · Tonsor (hairstyles) · realtime: Tiresias · Chronos · Natio · Cosmeta.
Website: https://cursedprograms.github.io/Changeling-Imago-Realtime/
Imago is Latin for a likeness, and the Romans' word for the wax masks cast from a person's face and kept in the family hall: a copy of a face that someone could wear. (It's also where "image" comes from.)
A live face swap for video calls. You become the face in personas/input.jpeg, with your own expressions
(your smile, your blinks, your mouth as you talk), and your call (Zoom, Teams, Discord, Meet) sees it as an
ordinary camera. The personas are AI-generated people who don't exist.
Each row above: a face as the camera sees it, the same frame through Changeling Imago Realtime, and the persona.
- Put the persona's face in
personas/input.jpeg(see below; the repo comes with one). - Start
run.bat(or.\run.ps1in PowerShell,./run.shin Git Bash or Linux). The first time it sets itself up: Python 3.11, PyTorch's CUDA libraries, InsightFace and the GPU ONNX runtime (about 3 GB), the swap model (models/inswapper_128.onnx, about 550 MB) and the occlusion model (models/xseg_1.onnx). Optional too: Sharper face's model,models/gfpgan_1.4.onnx(GFPGAN v1.4 as ONNX; also looked for inF:\AIModels\faceswap). Make it once fromGFPGANv1.4.pthwithscripts/export_gfpgan.py(run it with Changeling Imago'spsdenv, which has gfpgan). An NVIDIA GPU is needed for a usable frame rate. - A dialog asks first: use the default input (
personas/input.jpeg)? Yes for her; No opens a file picker inpersonas/to become someone else this time. - In your call, pick the camera OBS Virtual Camera (OBS Studio installed; it doesn't have to run).
The window (Changeling's look): the picture on the left; on the right the persona with Pick persona and Reload, switches for the swap, occlusion and the virtual camera, the smooth and strength sliders, the voice, and Record.
| Key | Does |
|---|---|
| S | swap on / off |
| O | occlusion on / off: hands, hair, a cup in front of your face stay in front of hers |
| H | Sharper face on / off: GFPGAN redraws the swapped face at 512 x 512 (the swap itself draws it at 128 x 128, which looks soft and blocky up close); about half the frame rate |
| V | virtual camera on / off |
| A | voice: off, female, male |
| E | Echo on / off: female and male in Changeling Echo's trained voices |
| C | record / stop: the swapped picture and your voice, to recordings/ |
| P | pick another persona: a file picker in personas/, no restart |
| R | read the persona's picture again (after replacing it) |
| Q or Esc | quit |
| Slider | Does |
|---|---|
| smooth % | steadies where the face sits from frame to frame (less jitter; too much and it lags) |
| strength % | how much of the swapped face shows: 100 is all of it |
| persona's skin % | how much of her skin you wear too: its tone and its fine detail (wrinkles, freckles, pores), which the swap alone leaves as yours. 0 by default; Changeling's realtime apps (Chronos, Natio, Tiresias Realtime) start it at 70. Your neck and ears keep your own tone, so a big change of skin tone shows at the jaw near 100 |
Female takes your voice to a woman's pitch (about 210 Hz), the formants (the resonance that sounds man or woman) up with it; male to a man's (about 115 Hz). Your own pitch is learned as you talk, so a deep voice and a light one both land in the right range. It's a phase vocoder on the CPU, about a tenth of a second behind. To reach your call it needs VB-Audio Virtual Cable (pick CABLE Output as the call's microphone); without it the voice still goes into recordings.
Echo (the switch under the voices, E, or --echo) speaks female and male in a trained voice
instead: Changeling Echo's echo_female and echo_male
(RVC). Your words and timing, someone else's voice, rather than yours moved up or down. It runs on the graphics
card in Echo's own environment (Echo's echo/live.py, a child process), a quarter second at a time: about a
third of a second behind you, under 100 ms of work per quarter second on a 6 GB card. It takes about
9 seconds to load the first time you pick a voice, and the phase vocoder speaks until it's ready. It needs
Changeling Echo next to this app (as in Changeling Studio) or in ECHO_DIR, with its voices trained.
Measured here: a man's voice at 158 Hz came out at 214 Hz as echo_female, and a woman's at 195 Hz came out
at 118 Hz as echo_male.
run_server.bat (or run_server.ps1, run_server.sh; it's run.bat --phone) uses your phone's camera and
microphone instead of a webcam, while this PC does the swap. The window shows a QR code: scan it with the phone
(on the same WiFi), accept the certificate warning once (the page makes its own certificate, because phones only
give a page the camera and microphone over HTTPS), and press Start. The phone sends its camera a picture at a
time and gets the swapped picture back on its screen; its microphone goes through the voice changer (and Echo). Out
to the call it's the same as a webcam: OBS Virtual Camera, the virtual cable, recordings. Tested here with a script
playing the phone: about 12 swapped pictures a second came back over the local network.
The page is on port 8443 (--phone-port changes it). Windows asks once whether Python may take connections on
your network: allow it on private networks. phone.py is the server, phone_page/ the page.
HD on the phone's page (on by default) sends and gets back pictures up to 1280 pixels at a high JPEG quality; off is 640 pixels (less WiFi, a little faster). The pictures can be upright or sideways: the virtual camera gets them fitted into its own size, with dark bars.
Record saves what the call would see and hear: the swapped picture and your voice (changed, or your
own with the voice off) as one .mp4 in recordings/, named by date and time. The picture is written at a
steady 30 fps in step with the sound; ffmpeg joins them at Stop. Click the line under the button to open
the folder.
run.bat --camera 1 (another webcam), run.bat --source clip.mp4 (a video or picture instead of the
webcam), run.bat --no-virtual (just the preview), run.bat --voice female --echo (start with a voice, in Echo's), run.bat --skin 70 (the persona's skin slider),
run.bat --mic "Focusrite" (pick the microphone by part of its name), run.bat --persona test2 (skip the question: a picture in personas/ by name, or any file), run.bat --source me.jpg --snapshot out.png (one
swapped frame, to try a persona).
The swap has two looks, a switch in the window: Classic (InsightFace's paste) or Deep-Live-Cam's look, written
anew after Deep-Live-Cam's ideas (dlc_style.py): a soft oval instead
of the square, the swapped face's colour matched to the skin it replaces, keep my mouth (your own lips, teeth and
tongue), a Poisson blend, opacity and sharpness, and smoothing between frames. Changeling Studio and Mimic (C++) have the same switch.
From the command line: --style deep-live-cam --keep-mouth 30.
Generated people who don't exist (AI Visual Workshop, Gemini...), not photos of real people: becoming a real person on a call isn't fair on them. What works best:
- looking straight at the camera, the whole face visible: no hands or hair over the eyes or mouth
- even light, sharp, a neutral face or a slight smile
- one face per image (if there are more, the biggest is used); a tight close-up is fine too
- The face: the persona's features and skin, your expressions and head movements.
- Not your hair, neck or body: a hood or a wig helps a lot. The voice is its own switch (above).
- Best facing the camera: sharp side profiles and hands over the face make it slip for a moment.
InsightFace finds the face in every frame (SCRFD, from the buffalo_l models in ~/.insightface), its
five key points are steadied against the last frame's, and inswapper_128 draws the persona's identity
(taken once from her image with InsightFace's recognition model) onto your face, all on the GPU through
onnxruntime-gpu. FaceFusion's XSeg occluder marks what's in front of your face (hands, hair, a cup) every
other frame, and the swap is left out there; when you blink, your own closed eyes show through. The
finished frame goes to the preview and to OBS's virtual camera.
InsightFace's models (buffalo_l, inswapper_128) are for non-commercial use only.



