Conform the face
The MetaHuman's face is conformed to the FaceNinja mesh, so the rig, LODs and facial animation all follow the new shape.
Product
Everything in a FaceNinja head.
What you download, how it is made, and what it does not do yet. No adjectives you cannot check.
MetaHuman topology, the person's own skin.
| Mesh | MetaHuman head topology with neck and upper shoulders, fixed vertex order and UVs, so every head lines up with every other |
|---|---|
| Eyes and teeth | Placed in their sockets; iris colour taken from the photo; a clear, reflective cornea over each eye that turns with the gaze |
| Skin maps | 4096 × 4096 base colour, normal, cavity, and specular / roughness |
| Wrinkles and pores | Relief from the photos goes into the normal map; freckles stay colour and do not become dents |
| Expressions | All 52 ARKit blendshapes, tongueOut included, refitted to the new face |
| Formats | GLB (with or without blendshapes), FBX, OBJ with textures, STL, PLY, MetaHuman ZIP |
| Download options | Maps as JPEG, PNG or none; normal maps in OpenGL or DirectX convention; 1K, 2K or 4K (GLB, OBJ, FBX). The MetaHuman ZIP is PNG, DirectX, 4K |
| Scale | Real units and life size (GLB: 1 unit = 10 cm), stated in every file; an estimate of the person's own size from the iris in the photo, applied only if you ask |
| Input | 1 to 15 photos of the same person |
| Time | About 5 to 10 minutes per head today |
Head hair is not part of the mesh: MetaHuman hair, brows and beards are grooms, and we recommend them separately (below). The skin under the hair is completed from the template, and beards are reduced to stubble in the skin so the beard groom has a clean surface to sit on.
A MetaHuman you can render, not a homework assignment. Early access
Importing a head into MetaHuman Creator normally means four manual jobs. Our Unreal automation does all of them.
Override the skin
Our base colour, normal and cavity maps replace the library skin. Without this step MetaHuman picks a stock skin and the person's freckles, moles and wrinkles are gone.
Put on the grooms
Hair, eyebrows, eyelashes, moustache and beard from the recommendation, with their colours set.
Match the body
A body type that joins the fitted neck and shoulders, so there is no seam and no mismatched collar line.
Early access. The automation runs in our own Unreal setup today and is being packaged for customers. Every plan already gets the MetaHuman ZIP: the head FBX in MetaHuman vertex order, the four 4K maps, the groom recipe and step-by-step Unreal instructions.
Hair, brows, lashes and beard, picked from the photo.
We segment the hair, brows and facial hair in your photos and measure them: length, how far the hair reaches, texture, coverage of the moustache, chin and jaw, how much is grey. A vision model names the style. Both are matched against the MetaHuman groom library.
- One recommendation per slot (hair, eyebrows, eyelashes, moustache, beard, peach fuzz) plus the next-best options with scores
- Melanin, Redness and Whiteness values for each, with colour swatches
- "Leave empty" when the person has no beard, instead of a default
groom.jsonfor tools andgroom.mdfor people, inside every MetaHuman ZIP
Hair_L_MessyClumps, Eyebrows_M_Shaded, Eyelashes_L_SlightCurl, no beard.Fitted to the person, not to the template.
Template
Fitted
Each 3D shape model gives us a neck. We measure its width, its front and back at mid-height, and where the trapezius sits, then:
- scale the template's neck to the measured width and depth,
- raise or lower the trapezius line to match,
- pull the front and sides of the neck onto the measured surface.
Long hair can hide the back of the neck, so the back only follows the measurement part of the way.
Check the head before you download it.
The Studio opens every finished head in a 3D viewer in the browser.
Expression presets
Smile, surprise, talk, frown and more, driven by the head's own ARKit shapes, to see that they moved with the face.
Hair preview
A preview of the person's hair, cut from the 3D shapes and coloured from the photos, to judge likeness. It is not exported: in Unreal the groom takes its place.
Groom recipe
The recommended grooms and colours next to the head, so you know what to pick in MetaHuman Creator before you open it.
Asynchronous, idempotent, boring in the good way.
| Step | What happens |
|---|---|
| Upload | Send photos with the create call, or upload them first and reuse their IDs |
| Create | One call starts a head; an Idempotency-Key makes retries safe |
| Wait | ?wait= long polling, a server-sent event stream, or a signed webhook |
| Request a file | Pick glb, fbx, obj, stl, ply or metahuman_zip; GLB and FBX with or without ARKit shapes; optionally texture format, normal-map convention, texture size, cornea and scale |
| Download | A short-lived signed URL; your API key never goes to the file host |
| Delete | Delete a head and its photos whenever you want |
Coming from KeenFace? The workflow is the same (create, wait, request a file, follow a signed link), so your integration keeps its shape; endpoints and field names differ. The docs include a migration guide.
API docsEvery intermediate image is checked, and redrawn if it fails.
- Same person. Face recognition compares the generated reference with your photos; too far apart and it is drawn again.
- Really neutral. The reference is checked for a frontal pose, a closed mouth and open eyes.
- Three opinions on shape. Three independent image-to-3D models each produce a head; fits that disagree with the others are dropped and the rest are averaged.
- A record of every run. Every check, measurement and warning is logged with the head, so a weak result can be traced to its cause.
Published method, published results.
We are running a benchmark on volunteers who gave written consent: the same photos go through FaceNinja and other services, and each head is compared with a reference scan (shape error in millimetres, by face region) and with the photos (face recognition similarity). We will publish the method, the results and the heads we are allowed to show, wins and losses alike.
Results are not in yet. This section will link to the full report when they are.
What FaceNinja does not do.
- Hair in the mesh. MetaHuman hair is a groom; we recommend it and show a preview, and the head itself is bald.
- Ear shape. Ears keep the template's shape.
- Pore-level skin from a scan. The texture comes from your photos at 4K; the finest grain on plain skin is the template's.
- Expression sets beyond the 52 ARKit shapes.
- Heads of people from a name or a web search. You upload photos of someone who agreed to it.
Later Reconstruction from calibrated multi-camera photo sets, with solved cameras and raw photo textures, is in research.
See it on your own face.
Your first head is on us: 1 free head when you sign up.