Style Waltz: Dancing Between Content and Style in Face Stylization

Yan Li, Zehao Li, Yizhu Jin, Lin Liu, XIAOPENG ZHANG, Wei Xue, Wenhan Luo, Qi Tian, Yike Guo

Keywords: Face Style Transfer, Point Interpolation

Abstract: Achieving precise artistic control while preserving identity remains a central challenge in facial image stylization, with most methods requiring costly training and offering limited flexibility. This paper introduces **StyleBrush**, a training-free stylization framework grounded in Riemannian geometry, which resolves this tension through a principled, dual-control optimization. Our core theoretical contribution is to reframe style transfer as a geodesic path-finding problem on a latent manifold. By leveraging the pullback metric, we establish a local isometry that validates optimizing a path’s energy in the embedding space as a means to approximate true geodesics, providing a rigorous foundation for style interpolation. This geometric framework is uniquely applied at two critical stages of the diffusion process: first, for interpolating content and style latents to ensure a semantically continuous fusion, and second, for modulating query features in self-attention layers to dynamically control stylization intensity. The unification of these two control mechanisms under a single geometric principle constitutes the primary novelty of our approach, enabling fine-grained and theoretically-grounded stylization control without any model training. Empirical validation on standard benchmarks confirms that our method significantly outperforms existing state-of-the-art approaches across a suite of quantitative and qualitative metrics.

Primary Area: generative models

Submission Number: 3399