[Feature Request] Add Implement of 'Kuramoto Orientation Diffusion'

Source code: https://github.com/KingJamesSong/OrientationDiffusion (the official implement use pure `torch`)

This paper introduces a nonlinear score-based generative framework that incorporates stochastic Kuramoto dynamics to model orientation-rich data. By formulating the forward process as synchronization and the reverse process as desynchronization, our method brings biologically inspired inductive biases into the diffusion process. Through wrapped Gaussian transitions and periodicityaware networks, the model naturally captures the geometry of angular data. Experiments show that our approach outperforms conventional baselines on orientation-dense datasets and remains competitive on general image generation tasks. This work highlights the potential of integrating biologically inspired synchronization dynamics as structured priors in generative modeling of orientation-dense data, paving the way for incorporating nonlinear dynamics into generative models.

Governing stochastic differential equations (SDEs) and representative image samples from our globally and locally coupled Kuramoto orientation diffusion models.

[60] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. ICLR, 2021.
[68] Jia Xue, Hang Zhang, and Kristin Dana. Deep texture manifold for ground terrain recognition. In CVPR, 2018.

---
```bibtex
@inproceedings{songKuramotoOrientationDiffusion2025a,
title = {Kuramoto {{Orientation Diffusion Models}}},
booktitle = {{{NeurIPS}}},
author = {Song, Yue and Keller, T. Anderson and Brodjian, Sevan and Miyato, Takeru and Yue, Yisong and Perona, Pietro and Welling, Max},
year = 2025,
url={\url{https://openreview.net/forum?id=dxK2QgEKvz}}
}
```

Commencez par examiner l’implémentation officielle liée à github.com/KingJamesSong/OrientationDiffusion ainsi que le Kuramoto Orientation Diffusion paper cité. Comparez ses composants PyTorch purs avec le dépôt diffusers afin de déterminer la surface d’intégration requise et l’approche de validation. La définition de « Done » n’est pas précisée dans l’issue ; le périmètre de l’implémentation et les tests attendus doivent être clarifiés par les mainteneurs.

Rédigé par le modèle d'indexation à partir du texte de l'issue.