Work

Dual-Identity Preservation in Text-to-Image Generation: A Unified Pipeline Integrating Latent Coupling, Cross-Attention Adaptation, and Artifact Mitigation

Year: 2025

Type: conference-paper

Abstract: This paper tackles the problem of maintaining dual identities in text-to-image generation, crucial for multi-concept image synthesis. It aims to overcome two main challenges: identity loss and partition boundary artifacts.The study presents a refined pipeline for dual-identity image generation. This method combines IP-Adapter with Latent Couple technology to preserve two distinct identities and uses a personalized LoRA model to reduce suture line artifacts. Extensive experiments show that this study's method outperforms existing techniques in identity preservation and image quality for both raw photo and style image generation.

DOI: 10.1145/3736426.3736480

Authors: Zhanfan Liang, Hai Huang, Xier Zhong, Haofan Liao

Institutions: City University of Macau, Zhuhai College of Science and Technology, University of Macau

Language: English

Topic: Generative Adversarial Networks and Image Synthesis

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

OpenAlex ID: W4412055685

FWCI: 0.785

Cites: 5

Cited by: 1

Related to: 10