Human Motion Unlearning

Edoardo De Matteis* · Matteo Migliarini* · Alessio Sampieri · Indro Spinelli · Fabio Galasso

*Equal contribution

Overview

We introduce Human Motion Unlearning, a novel approach to selectively remove specific motion concepts from text-to-motion generation models while preserving overall generation quality. Our work focuses on violence removal as a critical safety requirement, given that popular datasets like HumanML3D (7.7% violent content) and Motion-X (14.9% violent content) contain substantial amounts of aggressive behaviors.

Why Violence Unlearning?

Violence presents a unique challenge for unlearning because it spans from atomic gestures (e.g., a single punch) to highly compositional sequences. It demands fine-grained suppression without degrading non-violent sub-motions, providing a stringent benchmark for motion "forgetting" while addressing critical safety concerns in robotics and animation.

Our Approach: Latent Code Replacement (LCR)

We propose LCR, a training-free method that operates directly on the discrete latent space of VQ-VAE based models. By identifying violent codes through frequency analysis and replacing them with safe alternatives (plus noise for diversity), LCR executes in ~15 seconds—orders of magnitude faster than fine-tuning—while optimizing the trade-off between violence suppression and motion quality.

---

Qualitative Results

Violence Removal on HumanML3D

"A man does a run-up to kick something lying on the ground."

Before Unlearning (MoMask)	After Unlearning (LCR)

"A man stands up from the ground and then kicks with force."

Before Unlearning	After Unlearning (LCR)

"A man punches and then kicks the enemy."

Before Unlearning	After Unlearning (LCR)

You can check out more qualitative results on out website.

---

Quantitative Results

Violence Removal on HumanML3D

Performance on forget set (violent motions) and retain set (safe motions):

Forget Set - Lower FID and MM-Safe indicate successful violence suppression:

Method	FID →	MM-Safe ↓	Diversity →	R@1 →
MoMask D_r (Upper Bound)	16.36	4.50	6.96	0.118
MoMask (Original)	1.16	5.59	5.59	0.176
Fine-tuning	2.30	5.00	5.92	0.150
UCE	11.86	4.63	7.14	0.135
RECE	6.95	4.90	6.55	0.148
LCR (Ours)	15.66	4.77	6.00	0.125

Retain Set - Performance should match original model:

Method	FID ↓	MM-Dist ↓	Diversity →	R@1 ↑
MoMask D_r (Reference)	0.075	2.96	9.55	0.512
MoMask (Original)	0.041	2.93	9.63	0.520
Fine-tuning	0.070	3.03	9.68	0.501
UCE	0.090	3.10	9.73	0.497
RECE	0.144	3.12	9.81	0.493
LCR (Ours)	0.050	2.99	9.52	0.508

Violence Removal on Motion-X

Forget Set:

Method	FID →	MM-Safe ↓	Diversity →	R@1 →
MoMask D_r	9.94	10.43	17.19	0.172
MoMask (Original)	6.89	9.29	17.11	0.322
RECE	13.42	11.21	17.11	0.221
LCR (Ours)	7.08	9.36	17.17	0.317

Retain Set:

Method	FID ↓	MM-Dist ↓	Diversity →	R@1 ↑
MoMask D_r	11.66	9.03	19.87	0.321
MoMask (Original)	3.70	8.27	19.34	0.384
RECE	3.69	9.14	19.02	0.332
LCR (Ours)	3.66	8.33	19.34	0.381

↓ Lower is better, → Closer to original/reference is better, ↑ Higher is better

---

Key Features

• Comprehensive Benchmark: Includes filtered versions of HumanML3D and Motion-X with distinct forget/retain sets and standard evaluation metrics (FID, MM-Safe, R-Precision).
• Model Support: Compatible with discrete latent space models like MoMask and bidirectional autoregressive models like BAMM.
• Method Comparison: Benchmarks our LCR method against UCE, RECE, and Fine-tuning baselines.

Key Advantages of LCR

LCR is a training-free method that works directly on discrete latent codes, completing in just ~15 seconds. It offers the best trade-off between safety and quality, remaining robust against implicit prompting and "jailbreak" attempts without the need for expensive retraining.

---

Getting Started

Environment and Checkpoints

git clone --recurse-submodules https://github.com/Mamiglia/hmu.git
cd hmu

conda create -n momask python=3.8
conda activate momask
# Install requirements
pip install -r src/momask_codes/requirements.txt
pip install gdown --force-reinstall

# Download checkpoints
bash scripts/utils/prepare.sh
bash src/momask_codes/prepare/download_evaluator.sh
bash src/momask_codes/prepare/download_glove.sh

Dataset

HumanML3D - Follow the instruction in HumanML3D, then copy the result dataset to our repository:

cp -r ../HumanML3D/HumanML3D ./dataset/HumanML3D

Motion-X - Follow the instruction in Motion-X, then copy the result dataset to our repository:

cp -r ... ./dataset/Motion-X

Running

For running the experiments:

# 1. Split dataset into forget/retain sets
bash scripts/utils/split_dataset.sh --split_name violence --dataset HumanML3D --main_split train_val
bash scripts/utils/split_dataset.sh --split_name violence --dataset HumanML3D --main_split test
# 2. Apply LCR unlearning
bash scripts/eval/lcr.sh violence HumanML3D

Citation

If you find this work useful, please cite:

@inproceedings{dematteis2026hmu,
  title={Human Motion Unlearning},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  author={De Matteis, Edoardo and Migliarini, Matteo and Sampieri, Alessio and Spinelli, Indro and Galasso, Fabio}, year={2026}
}

---

Acknowledgements

This work builds upon several excellent open-source projects:

• MoMask for the text-to-motion generation framework
• BAMM for the bidirectional autoregressive motion model
• HumanML3D and Motion-X for the motion-language datasets

We thank the authors for making their code and data publicly available.

We acknowledge support from Panasonic, the PNRR MUR project PE0000013-FAIR, and HPC resources provided by CINECA.

---

Project Page | Paper

Repo Structure

├── assets
├── checkpoints
│   ├── HumanML3D -> t2m
│   ├── Motion-X
│   └── t2m 
├── dataset
│   ├── HumanML3D 
│   ├── __init__.py
│   └── Motion-X
├── glove
│   ├── our_vab_data.npy
│   ├── our_vab_idx.pkl
│   └── our_vab_words.pkl
├── README.md
├── scripts
└── src
    ├── bamm
    ├── eval
    ├── __init__.py
    ├── methods
    └── momask_codes