24 Feb 2024 | Chuan Guo1*, Yuxuan Mu1*, Xinxin Zuo2, Peng Dai2, Youliang Yan2, Juwei Lu2, Li Cheng1
The paper introduces a novel generative framework for 3D human motion stylization, leveraging the latent space of pretrained autoencoders to extract and infuse motion style. Unlike existing methods that operate directly in pose space, the proposed framework uses latent motion features (motion codes) to achieve more expressive and robust representation. The framework decomposes a motion code into a deterministic content code and a probabilistic style code, which is sampled from a prior Gaussian distribution. During training, the generator combines these components to reconstruct the corresponding motion codes, with cycle reconstruction and homo-style alignment techniques ensuring content preservation and style disentanglement. The approach supports various conditioning options, including deterministic stylization using reference motions, label-conditioned stylization, and unconditional stochastic stylization. Experimental results on three datasets demonstrate superior performance in style reenactment, content preservation, and generalization, outperforming state-of-the-art methods while being significantly faster. The key contributions include a novel generative framework, versatile stylization capabilities, and efficient runtime performance.The paper introduces a novel generative framework for 3D human motion stylization, leveraging the latent space of pretrained autoencoders to extract and infuse motion style. Unlike existing methods that operate directly in pose space, the proposed framework uses latent motion features (motion codes) to achieve more expressive and robust representation. The framework decomposes a motion code into a deterministic content code and a probabilistic style code, which is sampled from a prior Gaussian distribution. During training, the generator combines these components to reconstruct the corresponding motion codes, with cycle reconstruction and homo-style alignment techniques ensuring content preservation and style disentanglement. The approach supports various conditioning options, including deterministic stylization using reference motions, label-conditioned stylization, and unconditional stochastic stylization. Experimental results on three datasets demonstrate superior performance in style reenactment, content preservation, and generalization, outperforming state-of-the-art methods while being significantly faster. The key contributions include a novel generative framework, versatile stylization capabilities, and efficient runtime performance.