HyperAIHyperAI
2 months ago

EMDM: Efficient Motion Diffusion Model for Fast and High-Quality Motion Generation

Zhou, Wenyang ; Dou, Zhiyang ; Cao, Zeyu ; Liao, Zhouyingcheng ; Wang, Jingbo ; Wang, Wenjia ; Liu, Yuan ; Komura, Taku ; Wang, Wenping ; Liu, Lingjie
EMDM: Efficient Motion Diffusion Model for Fast and High-Quality Motion
  Generation
Abstract

We introduce Efficient Motion Diffusion Model (EMDM) for fast andhigh-quality human motion generation. Current state-of-the-art generativediffusion models have produced impressive results but struggle to achieve fastgeneration without sacrificing quality. On the one hand, previous works, likemotion latent diffusion, conduct diffusion within a latent space forefficiency, but learning such a latent space can be a non-trivial effort. Onthe other hand, accelerating generation by naively increasing the sampling stepsize, e.g., DDIM, often leads to quality degradation as it fails to approximatethe complex denoising distribution. To address these issues, we propose EMDM,which captures the complex distribution during multiple sampling steps in thediffusion model, allowing for much fewer sampling steps and significantacceleration in generation. This is achieved by a conditional denoisingdiffusion GAN to capture multimodal data distributions among arbitrary (andpotentially larger) step sizes conditioned on control signals, enablingfewer-step motion sampling with high fidelity and diversity. To minimizeundesired motion artifacts, geometric losses are imposed during networklearning. As a result, EMDM achieves real-time motion generation andsignificantly improves the efficiency of motion diffusion models compared toexisting methods while achieving high-quality motion generation. Our code willbe publicly available upon publication.

EMDM: Efficient Motion Diffusion Model for Fast and High-Quality Motion Generation | Latest Papers | HyperAI