Website: https://cursedprograms.github.io/Changeling-Cosmeta/
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.
A cosmeta was the servant who adorned a Roman lady: her makeup, her hair, her jewels (from the Greek kosmetes, the root of "cosmetics"). Changeling Cosmeta does the same for a photo. FaceApp-style cosmetics for photos: makeup and hair colour, painted onto the face itself so the light and texture underneath stay (a lipstick keeps the lips' lines, a hair colour keeps every strand). Two layers that stack, a basic editor, makeup and face paint copied from any photo, and photographic makeup looks or hair length through AetherCanvas AI. Beards (colour, stubble, shave, and the AI beards) moved to Changeling Tonsor, the barber, with the hairstyles. Every face above is an AI-generated person who doesn't exist.
The top row is the makeup; each row below is a person wearing it. The reference's face is warped onto hers triangle by triangle, so the drawn patterns land where they were: the clown's stars and tears, the skull's petals and web, the corpse paint's cracks, the festival dots. Face paint (the first five) takes the colours outright; Makeup (dark glam, last) takes only what the makeup changed and keeps her own skin tone. All the faces are AI-generated people who don't exist.
The face paint references are extreme on purpose: they're a stress test (every stroke has to land in the right place, on any face), not what Cosmeta is for. Everyday makeup, like the dark glam one, is the real use, and the subtler it is, the more natural it looks: compare the last column with the others.
Start run.bat (or .\run.ps1, ./run.sh). The first time it sets itself up: Python 3.11 with MediaPipe,
OpenCV, ONNX Runtime and customtkinter, and the face parser (models/face_parsing_int8.onnx, about 90 MB). No GPU
needed. run.bat photo.jpg opens with that photo.
| Control | Does |
|---|---|
| Open photo (Ctrl+O), Save (Ctrl+S) | the face is read once (a few seconds); Save writes the full-size result |
| Copy makeup from a photo | pick a photo of a face wearing makeup (makeup_refs/). Makeup: its lips, eyes, blush and brows come over, your own skin tone stays. Face paint: everything comes over, a clown's or a goth's white base too. It copies colours and style, not drawn patterns (cracks, stars, motifs) |
| Layer 1 / Layer 2 | two layers that stack, each with its own looks, editor, Opacity and Shown; Clear layer starts one again |
| Hair colour | 11 colours (platinum to electric blue) or any colour (+) |
| Lips | lipstick, with Gloss |
| Eyeshadow, Eyeliner, Blush, Brows | each with its own colours and strength |
| Blend | hair, lips, eyeshadow, blush and brows each choose how their colour goes on: Natural (Cosmeta's own), or Photoshop's Color (the classic for hair dye: your lightness, the colour's hue), Hue, Soft light, Overlay, Multiply, Screen |
| Smooth skin | an edge-keeping blur on the skin only |
| Editor | brightness, contrast, saturation, warmth, vignette, sharpen |
| AI looks (AetherCanvas AI) | makeup: natural, glam, smoky eyes, winged liner, red, nude or berry lips, glow, soft glam, gothic, bridal, freckles, no makeup; hair length: bangs, curly, straight, ponytail, bun, braids: sent to AetherCanvas AI when it's running (a minute or two each); Back to the original photo undoes them |
Every control answers at once: the preview is worked on a smaller copy, and Save paints the full-size photo.
-
Read the face, once. A SegFormer face parser (jonathandinu/face-parsing, int8 ONNX) labels every pixel: hair, skin, nose, lips, brows, eyes, ears, neck, clothes. MediaPipe Face Mesh finds the face's 468 points (for the eyelids and cheekbones).
-
Paint in LAB. Each look is a colour blended into its region: the region's brightness moves part of the way to the colour's, its tint all the way, its own fine variation stays. Gloss brightens where the lips already catch the light.
-
Facial hair is in Changeling Tonsor now. (
cosmetics/facial_hair.pystays for Changeling Cosmeta Realtime's live beard tools.) -
Copied makeup. Both faces are read with Face Mesh (478 points, plus a ring around each face reaching up the forehead). The reference is cut into those triangles and each one is warped onto yours, at full size, so every stroke lands where it was: a clown's star over your eye, a skull's teeth on your lips, whatever your face's shape. Face paint takes the reference's colours outright and keeps your skin's fine texture under them. Makeup takes only what the makeup changed, compared with the reference's own bare skin around it, and adds that to your skin. The lips take the reference's lip colour. Your eyes and open mouth stay your own (the reference's are filled in before warping). About a fifth of a second a photo; no model to download. Painted faces Face Mesh can't read straight away are cropped to the face OpenCV's YuNet finds and read again.
Model switches how it's copied: Warp (this, the default), GAN (classic) (the earlier model, DMT: Disentangled Makeup Transfer, colours and style at 256 pixels, if
models/dmt.onnxis on the PC) or Stable-Makeup (a diffusion model, about 8 GB of models inF:\AIModels\StableMakeup; it's offered once they're all there andsetup_stable_makeup.bathas set up its code and environment). One that isn't available says why.Stable-Makeup redraws a 512 x 512 square around the face (the identity and face shape held by two ControlNets), and Cosmeta puts it back on the face only. It runs in half precision and fits a 6 GB card (about 4.4 GB of models, 5.1 GB at its peak): about a minute to load the first time, then about 10 seconds a photo on an RTX 2060. Where Warp copies a pattern stroke for stroke, Stable-Makeup paints in the reference's style: a sugar skull keeps its white face, black nose and orange-and-purple eyes but not every petal and dot.
Its limits: a reference with an open mouth on a closed one (or the other way) leaves the paint around the lips less tidy. A reference turned well away from the camera stretches. Hats, noses and the like aren't 3D: a clown's red nose is painted on.
-
Two layers. Layer 2 is painted onto Layer 1's result with the same masks, each with its editor and opacity.
cosmetics/adjust.py (the editor) is the same file in every Changeling.
The code is MIT; The face parser was trained on CelebAMask-HQ, which is for non-commercial research; MediaPipe is Apache 2.0.





