GaitCrafter: Diffusion Model for Biometric Preserving Gait Synthesis

Sirshapan Mitra1, Yogesh S. Rawat1

1 Center for Research in Computer Vision, University of Central Florida

Paper (PDF) arXiv Code Back to Portfolio

Abstract

GaitCrafter is a diffusion-based framework for synthesizing realistic silhouette-domain gait sequences with identity preservation and controllable covariates (clothing, baggage, view). Incorporating synthetic samples improves recognition in low-label and privacy-constrained settings, and mixing identity tokens produces novel, consistent identities useful for training.

Method Diagram

GaitCrafter Method Diagram
Overview of the GaitCrafter framework: diffusion-based generation with controllable covariates and identity consistency.

Results

1) Generated Gait Sequences (30-frame cycles)

Animation
Temporally consistent sequences preserving gait structure across identities.

2) Controllable Covariates (View / Clothing / Baggage)

Animation Animation
Consistent identity with controlled viewpoint and attire/carrying conditions.

3) Quantitative Results (CASIA-B, 100%)

Table 1: Impact of synthetic data on recognition performancewith 100\% labeled data. $O$ means Original data, $O_D$ means diffusion generated original data and $O_N$ means diffusion generated novel data.
Novel IDs give larger gains than only increasing samples per existing ID.

4) Open-Set (20% IDs)

Table 2: Open-set with 20% IDs
Adding synthetic novel IDs improves performance, notably in CL condition.

5) Identity Preservation & Novel IDs (t-SNE)

t-SNE: real vs synthetic for known IDs t-SNE: known vs novel IDs
Synthetic samples co-cluster with real counterparts; mixed tokens yield distinct novel clusters.

6) Reverse Diffusion Process

Reverse diffusion process from noise to silhouettes
Gaussian noise → clean bimodal silhouette distribution across timesteps.

7) Multi-seed Variants per ID

Reverse diffusion process from noise to silhouettes
Different seeds introduce subtle motion variations while preserving identity.

BibTeX

@article{mitra2025gaitcrafter,
  title={GaitCrafter: Diffusion Model for Biometric Preserving Gait Synthesis},
  author={Mitra, Sirshapan and Rawat, Yogesh S},
  journal={arXiv e-prints},
  pages={arXiv--2508},
  year={2025}
}