I’m currently running into an issue when integrating ML Deformer with MetaHuman clothing, and I’m hoping to get some advice or technical insights from the community.
Environment & Pipeline Setup:
Engine Version: Unreal Engine 5.x
Skeleton: The clothing mesh and the MetaHuman body share the exact same skeleton hierarchy and bone definitions.
Body ML Deformer: My custom MetaHuman body has already been trained with ML Deformer. It works perfectly in Unreal Engine—when driven by the MetaHuman Control Rig, the mesh deformation responds accurately with zero issues. This confirms that our data generation pipeline, Maya-to-UE export workflow, and ML Deformer setup for the body are solid.
The Problem: I trained a separate ML Deformer asset for the clothing mesh using the same skeletal setup. However, after attaching the clothing to the MetaHuman and enabling its ML Deformer asset:
When driving the character using the Control Rig controllers, the clothing mesh experiences severe deformation artifacts (e.g., unexpected vertex offsets / clipping / mesh breakdown).
Disabling the clothing’s ML Deformer restores standard LBS (Linear Blend Skinning) behavior, so the base weight painting itself seems stable.
Questions:
Why does the ML Deformer respond correctly to Control Rig inputs on the body mesh, but fail on the clothing mesh despite using the same skeletal hierarchy?
Are there known caveats or best practices regarding ML Deformer when applied to clothing? (e.g., input bone list filtering, interaction with Leader Pose Component / Master Pose, or specific neural network training settings for loose geometry/cloth offsets?)
Any help or suggestions would be greatly appreciated! Thanks in advance.
The body mesh will deform correctly with its same skeletal hierarchy, however often the issue comes from how very different the deformation of the clothing is.
This could include loose offsets, separate wind deformation and overlapping geometries which are harder to predict/define using standard bone inputs/training parameters.
Before diving too deep into retraining, I’d love to narrow down a couple of things:
Which specific Unreal Engine 5 minor version are you currently on (e.g, 5.4, 5.5)?
When you trained the clothing ML Deformer asset, Did you include any cloth-specific bone curves or ensure that your Input Bone List in the ML Deformer Asset details panel fully accounts for the clothing’s unique twist/bend bones or are you sharing the exact same bone subset as the body?
While we look into those details, I’d suggest checking out a few things that usually help clear up clothing artifacts with ML Deformers:
Check the Input Bone List and Neighborhood Settings
Clothing meshes often flex independently of the underlying skeleton due to cloth simulation or loose offsets. If your clothing ML Deformer is using the exact same restricted bone list as the body, it might be missing crucial deformation influences.
You might want to expand the bone list or adjust the vertex neighborhood settings inside the ML Deformer Editor. You can read more about setting up inputs in the Epic Documentation on ML Deformer Setup
Evaluate Leader Pose Component Interactions
If your clothing uses a Leader Pose Component (or Master Pose Component) to follow the body skeletal mesh before the ML Deformer applies its corrective vertex deltas, check if the evaluation order is causing double-transform issues or conflicting offsets.
Let me know those engine version details and how your bone list is structured and we can look it closer.
Hi @BRS_GerDAngelo, thank you so much for the detailed response! Here are the exact details of my setup and what I’ve narrowed down:
1. Engine & Model Details:
Engine Version: Unreal Engine 5.8
ML Model: Neural Morph Model
Bone List: The clothing mesh has zero extra/custom cloth bones. It uses the exact same MetaHuman body skeleton hierarchy and shares the exact same Input Bone List as the body.
2. My Pipeline & Key Test Result (Crucial Finding):
In Maya, I transferred the skin weights directly from the MetaHuman body to the clothing mesh, generated the training animation using the official UE ML Deformer Maya plugin, exported the body ABC to Marvelous Designer for cloth simulation, and exported the simulated cloth ABC (ensuring identical rest pose and vertex topology).
The Smoking Gun: When I play the training Anim Sequence directly on the standalone Clothing Skeletal Mesh inside UE 5.8, the Neural Morph Model deforms the cloth 100% accurately—matching the MD simulation cache perfectly.
The Failure Point: The severe deformation artifacts only happen after attaching the clothing to the MetaHuman body and driving the body via Control Rig.
I also reproduced this exact same issue by adding a Control Rig to the clothing in Epic’s official ML Deformer Sample project.