Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HumanCM: One Step Human Motion Prediction

Liu Haojie, Gao Suixiang

TL;DR

This work tackles real-time 3D human motion prediction by moving beyond diffusion-based multi-step sampling to a one-step generation paradigm. By distilling diffusion dynamics into a self-consistent mapping in a latent DCT domain and integrating a Transformer-based spatiotemporal encoder with temporal embeddings, the approach preserves long-range temporal and inter-joint dependencies. The authors introduce a reconstruction-guided objective to stabilize training and demonstrate state-of-the-art or competitive accuracy on Human3.6M and HumanEva-I with inference speeds exceeding diffusion models by over two orders of magnitude. This advancement enables real-time, coherent motion forecasting suitable for interactive agents and AR/VR applications, with future work aimed at controllable and multi-agent motion synthesis.

Abstract

We present HumanCM, a one-step human motion prediction framework built upon consistency models. Instead of relying on multi-step denoising as in diffusion-based methods, HumanCM performs efficient single-step generation by learning a self-consistent mapping between noisy and clean motion states. The framework adopts a Transformer-based spatiotemporal architecture with temporal embeddings to model long-range dependencies and preserve motion coherence. Experiments on Human3.6M and HumanEva-I demonstrate that HumanCM achieves comparable or superior accuracy to state-of-the-art diffusion models while reducing inference steps by up to two orders of magnitude.

HumanCM: One Step Human Motion Prediction

TL;DR

This work tackles real-time 3D human motion prediction by moving beyond diffusion-based multi-step sampling to a one-step generation paradigm. By distilling diffusion dynamics into a self-consistent mapping in a latent DCT domain and integrating a Transformer-based spatiotemporal encoder with temporal embeddings, the approach preserves long-range temporal and inter-joint dependencies. The authors introduce a reconstruction-guided objective to stabilize training and demonstrate state-of-the-art or competitive accuracy on Human3.6M and HumanEva-I with inference speeds exceeding diffusion models by over two orders of magnitude. This advancement enables real-time, coherent motion forecasting suitable for interactive agents and AR/VR applications, with future work aimed at controllable and multi-agent motion synthesis.

Abstract

We present HumanCM, a one-step human motion prediction framework built upon consistency models. Instead of relying on multi-step denoising as in diffusion-based methods, HumanCM performs efficient single-step generation by learning a self-consistent mapping between noisy and clean motion states. The framework adopts a Transformer-based spatiotemporal architecture with temporal embeddings to model long-range dependencies and preserve motion coherence. Experiments on Human3.6M and HumanEva-I demonstrate that HumanCM achieves comparable or superior accuracy to state-of-the-art diffusion models while reducing inference steps by up to two orders of magnitude.
Paper Structure (13 sections, 9 equations, 3 figures, 2 tables, 1 algorithm)

This paper contains 13 sections, 9 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Human Motion Prediction
  4. Diffusion and Consistency Models
  5. Methodology
  6. Problem Definition
  7. Discrete Cosine Transform
  8. HumanCM
  9. Experiments
  10. Datasets and Metrics
  11. Implementation details
  12. Results
  13. Conclusion

Figures (3)

  • Figure 1: Comparison of generation time under identical hardware and batch size. HumanCM achieves over two orders of magnitude faster inference compared with diffusion-based models such as HumanMAC and TransFusion.
  • Figure 2: Overall training pipeline of HumanCM. The model learns a self-consistent mapping between noisy and clean motion representations under the consistency distillation framework.
  • Figure 3: Timeline comparison of motion generation processes. The blue box illustrates the traditional diffusion-based inference, where the model progressively denoises from $y(T)$ to $y(0)$ through $T$ iterative steps. In contrast, the green box highlights our HumanCM one-step inference, which directly maps the noisy input $y(T)$ to the clean motion state $y(0)$ in a single forward pass, achieving over two orders of magnitude faster generation.