SafeMo: Linguistically Grounded Unlearning for Trustworthy Text-to-Motion Generation

TL;DR

SafeMo tackles safety in diffusion-based text-to-motion by shifting from discrete codebook unlearning to continuous-space Minimal Motion Unlearning (MMU), guided by SafeMoEngine data synthesis. The framework absorbs harmful motion knowledge into a low-rank LoRA subspace and negates it at inference with class-aware scaling, preserving benign performance. It also introduces SafeMoVAE-29K and SafeMoVQ-29K, the first safe T2M datasets, and demonstrates superior forgetting of unsafe prompts while maintaining or improving safe-prompt fidelity across HumanML3D and Motion-X benchmarks. The approach yields strong safety-utility trade-offs, advancing trustworthy motion generation with a scalable, plug-and-play unlearning paradigm.

Abstract

Text-to-motion (T2M) generation with diffusion backbones achieves strong realism and alignment. Safety concerns in T2M methods have been raised in recent years; existing methods replace discrete VQ-VAE codebook entries to steer the model away from unsafe behaviors. However, discrete codebook replacement-based methods have two critical flaws: firstly, replacing codebook entries which are reused by benign prompts leads to drifts on everyday tasks, degrading the model's benign performance; secondly, discrete token-based methods introduce quantization and smoothness loss, resulting in artifacts and jerky transitions. Moreover, existing text-to-motion datasets naturally contain unsafe intents and corresponding motions, making them unsuitable for safety-driven machine learning. To address these challenges, we propose SafeMo, a trustworthy motion generative framework integrating Minimal Motion Unlearning (MMU), a two-stage machine unlearning strategy, enabling safe human motion generation in continuous space, preserving continuous kinematics without codebook loss and delivering strong safety-utility trade-offs compared to current baselines. Additionally, we present the first safe text-to-motion dataset SafeMoVAE-29K integrating rewritten safe text prompts and continuous refined motion for trustworthy human motion unlearning. Built upon DiP, SafeMo efficiently generates safe human motions with natural transitions. Experiments demonstrate effective unlearning performance of SafeMo by showing strengthened forgetting on unsafe prompts, reaching 2.5x and 14.4x higher forget-set FID on HumanML3D and Motion-X respectively, compared to the previous SOTA human motion unlearning method LCR, with benign performance on safe prompts being better or comparable. Code: https://github.com/AIGeeksGroup/SafeMo. Website: https://aigeeksgroup.github.io/SafeMo.

Figures (9)

• Figure 1: Discrete Motion Token vs. Continuous Motion Token.Discrete: generation is constrained by finite codebook entries, leading to quantization artifacts and piecewise transitions under the same prompt. Continuous: smoother kinematics and joint trajectories, natural phase transitions without staircase and jitter.
• Figure 2: Overview of the SafeMoEngine. We first classify and rewrite harmful texts (Level 2 & 3), route Level 1 texts to original motions, compose text conditions and syhthesize motions via two generative models, to construct SafeMoVAE-29K and SafeMoVQ-29K, respectively.
• Figure 3: Overview of SafeMo.Stage 1 (top): the unsafe stream optimizes through a harmful motion-specific loss and a random decoupling strategy, while the safe stream applies a negative preservation divergence. Only LoRA adapters on DiP are updated to obtain the pure harmful task vector. Stage 2 (bottom): we negate the learned harmful task vector via a motion-class aware $\alpha$, such that the model suppresses unsafe behaviors on unsafe prompts and preserve performance on safe prompts.
• Figure 4: Qualitative results of our models.
• Figure 5: Qualitative results of generated motions on unsafe prompts (Part I).