Manuscript received October 17, 2025; revised October 28, 2025; accepted November 21, 2025; published April 28, 2026.

Abstract—Facial nerve palsy results in impaired voluntary facial movements, and in Japan its severity is commonly assessed using the 40-point Yanagihara method. However, visual assessment is prone to inter-rater variability, and the use of actual patient images for educational purposes is restricted due to privacy concerns. To overcome these limitations, we propose a simulation framework that produces synthetic facial palsy images by leveraging a Stable Diffusion model with a partially fine-tuned ControlNet. Finetuning is confined to timesteps representing higher-level features, which allows the generation of pathological expressions while reducing the likelihood of reproducing patient-specific identity. Furthermore, a preprocessing stage modifies the size and placement of facial components and incorporates contour information from other faces. This design helps keep the fidelity of simulated expressions while maintaining strong privacy protection. The proposed method offers a practical resource for medical training and clinical research, where realistic and privacy-preserving facial data are required.

Keywords—facial palsy, diffusion models, ControlNet, finetuning, privacy preservation

Cite: Atsushi Tajima, Masataka Seo, and Yen-Wei Chen, "Privacy-Preserving Simulation of Facial Palsy Expressions Using Diffusion Models," Journal of Image and Graphics, Vol. 14, No. 2, pp. 294-302, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.