Identity-Preserved Portrait Generation for Fictional Worlds via Visual Reference Guidance

Abstract

Identity-preserving portrait generation has gained significant popularity in recent years due to its success in realistic face synthesis; however, it still struggles to balance biometric fidelity and stylistic integrity within fictional world applications. Furthermore, the heavy reliance on text prompts remains a critical limitation for generating unique artistic styles. This study proposes a hybrid approach that ensures identity preservation via the InstantID model while simultaneously learning stylistic features directly from visual references by disentangling content and style through B-LoRA. Experimental findings and quantitative metrics demonstrate that the developed approach achieves high CLIP-I (0.739) and CLIP-T (0.270) scores in fictional world scenarios, while maintaining a competitive level of identity preservation performance compared to current state-of-the-art models in the literature.