Causal Motion Tokenizer for Streaming Motion Generation

Biao Jiang, Xin Chen, Ailing Zeng, Xinru Sun, Fukun Yin, Xianfang Zeng, Xuanyang Zhang, Gang Yu, Tao Chen; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 2024-2034

Recent advancements in human motion generation have leveraged various multimodal inputs, including text, music, and audio. Despite significant progress, the challenge of generating human motion in a streaming context--particularly from text--remains underexplored. Traditional methods often rely on temporal modalities, leaving text-based motion generation with limited capabilities, especially regarding seamless transitions and low latency. In this work, we introduce MotionStream, a pioneering motion-streaming pipeline designed to continuously generate human motion sequences that adhere to the semantic constraints of input text. Our approach utilizes a Causal Motion Tokenizer, built on residual vector quantized variational autoencoder (RVQ-VAE) with causal convolution, to enhance long sequence handling and ensure smooth transitions between motion segments. Furthermore, we employ a Masked Transformer and Residual Transformer to generate motion tokens efficiently. Extensive experiments validate that MotionStream not only achieves state-of-the-art performance in motion composition but also maintains real-time generation capabilities with significantly reduced latency. We highlight the versatility of MotionStream through a story-to-motion application, demonstrating its potential for robotic control, animation, and gaming.