Abstract
Traditional text-to-motion frameworks often struggle with "identity bias," producing generic motions that fail to account for the physical attributes of the subject. We present AttrMoGen, a novel framework that decouples action semantics from physical attributes like age, gender, height, and weight. By employing a causal decoupling strategy and a new large-scale dataset, HumanAttr, AttrMoGen generates realistic, attribute-specific 3D human motions from simple textual prompts.
Introduction
While text-to-motion AI has advanced significantly, existing models typically generate "average" motions. A toddler’s walk, an athlete’s sprint, and an elderly person’s stride are biomechanically distinct, yet current models often conflate these identities with the action itself. This "semantic coupling" prevents fine-grained control over the generated output. Furthermore, the field suffers from a lack of diverse datasets annotated with physical attributes.
Methodology
AttrMoGen addresses these challenges through a two-stage architecture designed for disentanglement.
Decoup-VQVAE
The core of the system is the Decoup-VQVAE, which separates "what" an action is from "who" is performing it. Inspired by Structural Causal Models (SCM), the encoder utilizes a Causal Information Bottleneck (CIB) to filter out attribute-related information, producing "attribute-free" semantic tokens. The decoder then reconstructs the motion by combining these clean tokens with specific attribute labels. To enforce this separation, a proxy attribute classifier is used during training; the system is optimized such that the classifier cannot determine the subject's attributes from the semantic tokens alone.
Semantics Generative Transformer
A Transformer model, based on the MoMask architecture, is trained to predict these attribute-free tokens from textual prompts. During inference, the user provides both a text prompt (e.g., "walking") and a set of target attributes (e.g., "Male, Age 60-80"). The model then synthesizes a motion that aligns with both the semantic intent and the physical identity.
The HumanAttr Dataset
To support this research, the authors introduced HumanAttr, a comprehensive dataset compiled from nine existing sources. It contains 18.2k motion sequences from 640 subjects, with ages ranging from 5 to 88 years. It provides gender labels for all sequences and height/weight data for 74% of the entries, offering the diversity needed for attribute-aware training.
Results and Discussion
AttrMoGen demonstrates superior performance on the HumanAttr benchmark compared to state-of-the-art models like MoMask and MDM.
- Motion Quality: The model achieved a Frechet Inception Distance (FID) of 0.089, significantly outperforming MoMask's 0.245.
- Text Alignment: Removing attribute "noise" improved R-Precision (Top-1) from 0.685 to 0.705, indicating that disentanglement helps the model better understand textual semantics.
- Attribute Accuracy: An independent classifier identified the correct gender in generated motions with 99% accuracy, proving the model successfully encodes physical traits into the movement style.
Conclusion
AttrMoGen represents a significant step toward personalized human motion synthesis. By treating physical attributes as non-causal factors for action semantics, the framework allows for the generation of diverse, realistic motions tailored to specific identities. Future work will focus on improving the granularity of height and weight controls and integrating physics-based constraints to ensure physical plausibility across all body types.
Summary updated