Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unconditional Human Motion and Shape Generation via Balanced Score-Based Diffusion

David Björkstrand, Tiesheng Wang, Lars Bretzner, Josephine Sullivan

TL;DR

This work shows that unconditional human motion and shape generation can match state-of-the-art diffusion models without auxiliary losses or over-parameterized inputs by applying principled feature normalization, per-group and per-time-step uncertainty weighting, and a minimal SMPL-based representation. Building on EDM and EDM2, the method leverages a structure-preserving SMPL parameterization, rotation-aware normalization, and magnitude-based standardization to maintain balanced training dynamics, achieving competitive FID and diversity with as few as 31 neural function evaluations. The approach enables direct joint generation of motion and shape, supports probability-flow sampling via PF-ODE, and demonstrates that careful theoretical motivation and ablations yield tangible performance gains while simplifying the training pipeline. The results suggest broad applicability to conditional generation and other domains with heterogeneous feature types, highlighting a scalable recipe for balanced diffusion in complex, multi-group data settings.

Abstract

Recent work has explored a range of model families for human motion generation, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and diffusion-based models. Despite their differences, many methods rely on over-parameterized input features and auxiliary losses to improve empirical results. These strategies should not be strictly necessary for diffusion models to match the human motion distribution. We show that on par with state-of-the-art results in unconditional human motion generation are achievable with a score-based diffusion model using only careful feature-space normalization and analytically derived weightings for the standard L2 score-matching loss, while generating both motion and shape directly, thereby avoiding slow post hoc shape recovery from joints. We build the method step by step, with a clear theoretical motivation for each component, and provide targeted ablations demonstrating the effectiveness of each proposed addition in isolation.

Unconditional Human Motion and Shape Generation via Balanced Score-Based Diffusion

TL;DR

This work shows that unconditional human motion and shape generation can match state-of-the-art diffusion models without auxiliary losses or over-parameterized inputs by applying principled feature normalization, per-group and per-time-step uncertainty weighting, and a minimal SMPL-based representation. Building on EDM and EDM2, the method leverages a structure-preserving SMPL parameterization, rotation-aware normalization, and magnitude-based standardization to maintain balanced training dynamics, achieving competitive FID and diversity with as few as 31 neural function evaluations. The approach enables direct joint generation of motion and shape, supports probability-flow sampling via PF-ODE, and demonstrates that careful theoretical motivation and ablations yield tangible performance gains while simplifying the training pipeline. The results suggest broad applicability to conditional generation and other domains with heterogeneous feature types, highlighting a scalable recipe for balanced diffusion in complex, multi-group data settings.

Abstract

Recent work has explored a range of model families for human motion generation, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and diffusion-based models. Despite their differences, many methods rely on over-parameterized input features and auxiliary losses to improve empirical results. These strategies should not be strictly necessary for diffusion models to match the human motion distribution. We show that on par with state-of-the-art results in unconditional human motion generation are achievable with a score-based diffusion model using only careful feature-space normalization and analytically derived weightings for the standard L2 score-matching loss, while generating both motion and shape directly, thereby avoiding slow post hoc shape recovery from joints. We build the method step by step, with a clear theoretical motivation for each component, and provide targeted ablations demonstrating the effectiveness of each proposed addition in isolation.
Paper Structure (32 sections, 40 equations, 3 figures, 3 tables)

This paper contains 32 sections, 40 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Preliminaries: EDM & EDM2
  4. The EDM score-based generative method
  5. EDM2 and tools for standardization
  6. Steps towards principled human motion and shape generation
  7. Motion representation
  8. Baseline generative method
  9. Input feature normalization using expected magnitude
  10. Gradient analysis of uncertainty weighting
  11. Per feature group uncertainty weighting
  12. Addressing feature group dimensionality
  13. Experiments
  14. Dataset and evaluation metrics
  15. Comparison to previous works
  16. ...and 17 more sections

Figures (3)

  • Figure 1: Unconditionally generated samples from our final model. SMPL parameters are generated directly and the mesh is extracted with the SMPL-H model. Generated with 31 NFEs. Darker color indicates later frames in the sequence. See supplementary videos for more qualitative results.
  • Figure 2: Average L2 norms of gradients with respect to $F_\theta$ over diffusion time steps at different points during training. From epoch 100 (lightest) to epoch 600 (darkest). Calculated using PyTorch autograd, with a batch size of 32 and averaged over entire training dataset. (a) Gradient norms before and after re-balancing. (b) Gradient norms per feature group before per feature group balancing. (c) Gradient norms per feature group after per feature group balancing.
  • Figure 3: PF-ODE evaluation on the full-length test motions. (a) NLL in unormalized feature space vs. NFEs. (b) Round-trip error in normalized feature space vs. backwards NFEs. (c) Round trip error in normalized feature space vs. backwards $\rho$.