Generating Consistent Full-Body Avatars: Stratified Motion Diffusion for Decoupled Kinematics

Table of Links

Abstract and 1. Introduction

2. Related Work

   2.1. Motion Reconstruction from Sparse Input

   2.2. Human Motion Generation

3. SAGE: Stratified Avatar Generation and 3.1. Problem Statement and Notation

   3.2. Disentangled Motion Representation

   3.3. Stratified Motion Diffusion

   3.4. Implementation Details

4. Experiments and Evaluation Metrics

   4.1. Dataset and Evaluation Metrics

   4.2. Quantitative and Qualitative Results

   4.3. Ablation Study

5. Conclusion and References

Supplementary Material

A. Extra Ablation Studies

B. Implementation Details

3.3. Stratified Motion Diffusion

After encoding and expressing different human motions as latents, we aim to properly sample from the latent space for full-body motion reconstructions and match the sparse observations.

Although disentangling the full-body motions into upper and lower parts enhances effectiveness and efficiency for motion representation learning, it’s crucial to include the correlation between two body parts during generation. Otherwise, severe inconsistency would be witnessed in reconstructed full-body motions. To this end, we propose Stratified Motion Diffusion to sample upper-body and lowerbody latent in a cascaded manner with explicit considerations of the correlations mentioned above.