[Feature Request] Light-weight inference pipeline tailored for driving standard FBX/Humanoid avatars (Skip EHM mesh reconstruction)

[SABBA-HOSTUMEX]

Hi PEAR Authors,

First of all, thank you for this incredible work! The pixel-aligned accuracy and the low-latency design of PEAR are absolutely amazing.

I am currently working on a real-time motion capture project using PEAR to drive standard 3D avatars (e.g., standard FBX models, VRM, or Unity Humanoid hierarchy) in a game engine via UDP.

Since my goal is strictly to animate an existing production-ready FBX character, I only need the joint rotation data (such as the Axis-Angle representations for Body, Left Hand, and Right Hand). I do not need the dense 3D mesh vertices, nor do I need to run the full ehm() forward kinematics / linear blend skinning (LBS) on the Python side, which introduces significant computational overhead (pushing my RTX 3090 "3D" engine usage to 95% at times).

Is it possible to provide a streamlined, inference-only pipeline that:

1. Only extracts the raw skeletal structure and local joint rotations (e.g., Zone A, B, and C).
2. Completely bypasses the EHM mesh reconstruction and the generation of the 358-float vector (skipping Zone D, E, F).
3. Provides a direct mapping concept from PEAR's predicted joint rotations to standard skeletal hierarchies?

Having a lightweight pipeline dedicated solely to extracting rotation data would make PEAR the ultimate tool for real-time XR and game development.

Thank you again for your time and contribution to the community!