MotionGen (Dissertation)

Text in, AnimationClip out

MotionGen is a Unity editor plugin for controllable, interactive 3D human motion generation. You type a prompt, and the result is written straight into the project as a humanoid AnimationClip asset. It supports single and batched generation, text-prompted regeneration of a single segment, inbetweening between two clips, and multi-segment composition.

Three models, no cloud

A local Python backend serves three text-to-motion models, T2M-GPT, MoMask and MDM, and Unity talks to it over gRPC. There is no cloud call and no external account anywhere in the loop. Running locally was the constraint the rest of the design had to work around, the planner included.

There is no cloud call and no external account anywhere in the loop.

Long prompts get a plan

The dissertation is titled 'Beyond Simple Prompts'. A local Gemma model takes a complex prompt and splits it into a per-segment JSON plan: which model runs each segment, how long it lasts, how it transitions, and what it is anchored to in the scene. The segments then compose into one clip.

Ranking the bad takes

Generative motion produces takes that are technically valid and obviously wrong. The variant ranker puts a number on that: it scores clips on foot skating, jerk, root drift and ground penetration, and orders them so the best one comes first. The ranker was calibrated as one of the four evaluation methods rather than assumed to work.

How it was evaluated

Four methods: a comparative within-subjects user study, an internal model-fit pre-screen, a latency benchmark for the local planner, and a calibration of the variant ranker. In the user study MotionGen beat Unity's own AI tooling on all four measures, SUS, NASA-TLX workload, task time and ease, with large effect sizes. The dissertation was graded 77.

The code is public and MIT licensed, and the dissertation PDF is on this site. The Unity Asset Store release is not out: it is pending University approval, and until that clears the plugin is something you clone rather than install.