ML Deformer and Meta Human Facial Expressions

Unreal Engine has a machine-learning deformer - ML Deformer - that learns complex muscle and tissue deformation from training data and reproduces it in real time. It is mostly discussed for bodies; its potential for facial animation is largely unexplored. This post looks at an experiment that applies ML Deformer to MetaHuman faces, with Mesh Morpher preparing the training data entirely inside Unreal Engine.

The experiment: FACS data meets ML Deformer

Elvis Morelli trained the ML Deformer using FACS (Facial Action Coding System) scan data from 3dscanstore.com, using Mesh Morpher to prepare the training data in-engine - no external DCC in the loop. The results are impressive:

▶ Watch the video on YouTube

The video demonstrates custom FACS data driving MetaHuman characters - a path to noticeably more realistic facial deformation than the stock rig alone.

The enabling piece: sculpting animation data in-engine

Training an ML Deformer needs paired data: poses and the corrected shapes they should produce. That is exactly what Mesh Morpher Graph makes editable in-engine - it can sculpt Geometry Caches and Animation Sequences directly, through the node interface. This tutorial shows geometry cache sculpting:

▶ Watch the video on YouTube

And this one walks through sculpting animation sequences step by step:

▶ Watch the video on YouTube

Why this matters

The bottleneck for ML-driven facial deformation has never been the runtime - it is authoring the training data. When the corrections can be sculpted on the actual animation, in the engine, with Delta Mush and the GPU brushes available for cleanup, the iteration loop tightens from days to hours. For teams already customizing MetaHumans with Mesh Morpher, ML Deformer becomes a natural next step up in facial fidelity.