SCIENCE@home

Dynamic Feature Fusion for Structure-Aware Portrait Stylization and Perception-Guided Aesthetic Evaluation

Wenjing Qi,Chuanyan Zhang

Published 2025 in IEEE Access

ABSTRACT

Portrait stylization aims to transform photographic portraits into compelling artworks while preserving identity and structural integrity. Existing methods often suffer from limited adaptability to local facial regions, insufficient multi-style fusion, and reliance on subjective evaluation criteria. To address these limitations, we propose LrNStyle++, a structure-aware portrait stylization framework that incorporates Dynamic Feature Fusion (DFF) across multiple stages of the generative process. A key innovation of our approach is the introduction of the Comprehensive Aesthetic Evaluation (CAE) module, which leverages deep neural features to align stylization outputs with human perceptual preferences. The CAE module employs a dynamic feature fusion mechanism built upon a feature-level gated attention network, enabling adaptive integration of content and style representations. This perceptually guided design establishes a unified framework that enhances the realism, coherence, and aesthetic quality of stylized portraits. Experimental results demonstrate that LrNStyle++ outperforms existing methods in terms of content preservation, stylization quality, and computational efficiency, achieving real-time inference even on mobile devices. Our work makes significant strides in both the technical and perceptual dimensions of portrait stylization, offering a new perspective on multi-style fusion and perceptual alignment.

PUBLICATION RECORD

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-36 of 36 references · Page 1 of 1

CITED BY

  • No citing papers are available for this paper.

Showing 0-0 of 0 citing papers · Page 1 of 1