PHOTO TO METAHUMAN

Our Lapu-Lapu game needs faces people recognise, and a small studio cannot afford a scanning rig or a character team for every one. Unreal's MetaHuman tools can build a production-ready digital human, but the fast route assumes an iPhone capture of the real person in the room.

So we set a test. Could a state-of-the-art AI model work as game-dev crew: read the engine's Python API, find a route from a single still photo to a finished MetaHuman, and recover from its own mistakes? Two models got the same brief on the same project, Claude Opus 5.5 and Claude Fable 5.1. Both ran on one 16 GB Mac and shared one editor. This page follows both runs: Fable 5.1 drove Epic's own identity solver from the photos, and further down, Opus 5.5 measured the faces first with research models.

The subjects were Apo Whang-Od, Michelle Dee and Sara Duterte, chosen because good public-domain or Creative Commons photos of them exist. These are technical likeness studies. They are rough, and none of the three had any part in them.

Be clear about what this shows. The pipeline is complete: every stage runs from a script and ends in an asset a game can use. The likeness is not there yet. Bone structure and proportions carry over, age and expression mostly do not, and one source portrait was smiling, so the smile is baked into the neutral face.

The Opus 5.5 run took a different road to the same three faces. Instead of estimating depth and handing photos straight to Epic's solver, it measured the faces first. A research model reads where each pixel of a face sits on a 3D head; every photo of a person is combined into one shape; only then does the shape go to Epic's tools. It also went further on the look: hair, brows, makeup and the lit portrait set at the top of this page.

Same faces, two routes. Skin and hair can flatter a face, so here are the bare shapes from both runs, rendered with the same clay material and framing.

The identity solve keeps lumps, a baked-in smile and a distorted forehead; the measured heads keep each face's proportions and stay clean enough to rig. Measure, don't generate: a likeness has to come from the person, not from a model's idea of a face.

Where the likeness still falls short. Measuring fixes proportions, not everything. Whang-Od's great age shows more in the skin texture than in the shape, because the 3D face model cannot fold skin the way more than a century does. Michelle Dee's photos are small and nearly all smiling, which limits how much of her face could be measured. And both Sara Duterte and Michelle Dee still carry a reddish rim under the eyes from the skin settings, which needs an artist's pass.

One more lesson came from the sources. A Commons photo filed under Whang-Od's village turned out to show another Kalinga woman. It was caught before publishing; it carried a small weight in the fit, and removing it moves her face shape by less than 0.3 mm.

The built characters went into their own level with the LOVE wordmark of the Philippine tourism brand and our logo as emissive neon behind them. Three camera moves were authored as Level Sequences and rendered at 1080p.

None of this was smooth, and the failures are the useful part.

The engine crashes on honest mistakes. Capture data without a depth sequence, a camera named anything other than "Video" and "Depth", or a build started while stale preview actors exist: each one kills the editor outright. The model read the crash stack each time, found the cause and wrote the rule down so it was not repeated.

What you see is not always what you built. The preview actor ships with a debug material that paints coloured regions and a wireframe over the face. For a while the renders looked broken when the textures underneath were fine.

Memory is the real limit on a laptop. Three cinematic MetaHumans rendered through the editor pushed it to 50 GB on a 16 GB machine. The fix was to quit the editor and render in a standalone process, which then crashed on a project startup script that assumed the editor was open.

Not every plugin runs everywhere. One store plugin shipped Windows binaries only and was recompiled for Mac after a two-line patch. Another was built against a different engine release and stays switched off.

A model cannot fix a bad source. A smiling portrait gives a smiling neutral face. Small faces, under about 150 pixels wide, track badly. In the Fable 5.1 run, hair, brows and clothing are still to do, so its heads are bald.

Next on the list: clothing from the MetaHuman wardrobe, a neutral source photo to replace the smiling one, the skin fixes listed above for the measured faces, the same faces through Epic's web Creator for comparison, and then into the Lapu-Lapu game as characters you can meet.