DeMoGen: Decomposing Human Motion with Diffusion Models
Analysis
This paper introduces DeMoGen, a novel approach to human motion generation that focuses on decomposing complex motions into simpler, reusable components. This is a significant departure from existing methods that primarily focus on forward modeling. The use of an energy-based diffusion model allows for the discovery of motion primitives without requiring ground-truth decomposition, and the proposed training variants further encourage a compositional understanding of motion. The ability to recombine these primitives for novel motion generation is a key contribution, potentially leading to more flexible and diverse motion synthesis. The creation of a text-decomposed dataset is also a valuable contribution to the field.
Key Takeaways
- •Proposes DeMoGen, a decompositional approach to human motion generation.
- •Employs an energy-based diffusion model for learning motion primitives.
- •Introduces three training variants to encourage compositional understanding.
- •Demonstrates the ability to recombine primitives for novel motion generation.
- •Constructs a text-decomposed dataset to support compositional training.
“DeMoGen's ability to disentangle reusable motion primitives from complex motion sequences and recombine them to generate diverse and novel motions.”