Project Page

Frequency-Enhanced Diffusion Models:
Curriculum-Guided Semantic Alignment for Zero-Shot Skeleton Action Recognition

Frequency-aware diffusion restores fine-grained motion dynamics that standard DiT backbones oversmooth in zero-shot skeleton action recognition.

Yuxi Zhou1,*, Zhengbo Zhang2,*, Jingyu Pan3, Zhiyu Lin4, Zhigang Tu1,5
1Wuhan University 2Singapore University of Technology and Design 3School of Geodesy and Geomatics, Wuhan University 4School of Mathematics and Statistics, Wuhan University 5Wuhan University Shenzhen Research Institute
*Equal contribution
Code BibTeX

Abstract

Zero-Shot Skeleton Action Recognition relies on semantic guidance to recognize unseen actions, but diffusion-based approaches still suffer from spectral bias: they preserve coarse pose structure while oversmoothing the high-frequency dynamics that distinguish fine-grained actions. FDSM addresses this issue with three coordinated components. A Semantic-Guided Spectral Residual Module restores discriminative high-frequency content in the latent space. A Timestep-Adaptive Spectral Loss aligns spectral supervision with the coarse-to-fine denoising trajectory. A Curriculum-based Semantic Abstraction strategy transfers rich LLM-generated motion descriptions into sparse-label inference. Across NTU RGB+D, PKU-MMD, and Kinetics-skeleton benchmarks, FDSM consistently improves over prior state of the art and produces more faithful motion representations.

Method

Method overview.

Pipeline of the FDSM framework across training and inference phases

Results

Main benchmark results.

Main benchmark results table for FDSM on NTU-60 and NTU-120

BibTeX

@article{zhou2026fdsm,
  title   = {Frequency-Enhanced Diffusion Models: Curriculum-Guided Semantic Alignment for Zero-Shot Skeleton Action Recognition},
  author  = {Yuxi Zhou and Zhengbo Zhang and Jingyu Pan and Zhiyu Lin and Zhigang Tu},
  year    = {2026},
  note    = {Project page version. Please update with the final publication metadata.},
  url     = {https://github.com/yuzhi535/FDSM}
}