Documentation

Quick start — 2 minutes, it already works

The plugin ships with 37 motion clips built in: no server, no account, no internet, nothing to configure.

1. Install the plugin (via Fab / the Plugins window) and restart the editor.
2. Open Tools → Kinisis Motion and click Add Demo Character.
3. Type a prompt (or pick one from Examples) and press Generate — the character plays it right in the viewport. Bake it, retarget it, grade it with Auto-TD: the whole pipeline works offline out of the box.

Upgrade to live generation (any prompt)

When you want motions the built-in library doesn't have, connect a motion server — three options:

• Your own NVIDIA GPU (6 GB+ VRAM): one click. In server mode press Install Local Server in the panel (~11 GB one-time download from official sources, no accounts). When it finishes, the plugin offers to switch modes, then starts and stops the server for you in the background (http://127.0.0.1:9090).
• A rented cloud pod (+ API key): Extras/pod_bootstrap.sh sets one up in a single command (~$0.40/hr with typical providers).
• No GPU? Run the included offline demo server: python Extras/kimodo_server_demo.py

The pipeline

Step	What happens
Generate	Your prompt → HTTP → motion server → BVH clip back in seconds
Steer	Optional: set Walk Target Actor + Walk To Target — "walk to the box" ends at your box; Land On Target jumps/climbs onto it
Retarget	Built-in C++ retarget onto Mannequin/MetaHuman — no IK Rig setup needed; ground conform + foot clamp included
Bake	Select the actor → Bake Last Motion To Sequence → a normal UAnimSequence in /Game/KinisisMotions. From there, retarget to ANY character with UE's own IK Retargeter
Validate	Auto-TD checks every clip: foot sliding, loop seams, joint limits, ground contact, jitter — flagged frames flash red in the review HUD

Auto-TD: validate your motion capture and animations

Not just for generated motion — point it at any BVH (Xsens, Rokoko, OptiTrack, hand-keyed):

pip install numpy
python Extras/autotd/mocap_validator.py your_clip.bvh --out-dir reports   # verdict + flagged frames
python Extras/autotd/mocap_autofix.py  your_clip.bvh --out-dir reports    # safe repairs, source untouched

• No black box — five deterministic checks, pure math. Same clip in, same verdict out.
• Your source is sacred — reports and repaired copies are written next to your clip; the original is never modified.
• Artists stay in charge — creative calls (e.g. foot-plant timing) are flagged, not "fixed".

Auto-TD needs no GPU and no server — any machine with Python 3.9+ and numpy.

Requirements at a glance

Tier	Hardware	Internet	Accounts
Built-in clips	Any machine	None — fully offline	None
Live local generation	NVIDIA GPU 6 GB+ VRAM, ~15 GB disk (16 GB RAM recommended under 10 GB VRAM)	One-time ~11 GB download, then offline	None
Cloud pod	Any machine	Connection to your pod	Pod provider only

Advanced full-precision mode: ~17 GB VRAM or ~16 GB free system RAM (32 GB machine recommended); free Hugging Face account required. See the included SERVER_SETUP.md.