AnyMS: Training-Free Multi-Subject Customization with Layout Guidance

This paper introduces AnyMS, a novel training-free framework for multi-subject image synthesis. It addresses the challenges of text alignment, subject identity preservation, and layout control by using a bottom-up dual-level attention decoupling mechanism. The key innovation is the ability to achieve high-quality results without requiring additional training, making it more scalable and efficient than existing methods. The use of pre-trained image adapters further enhances its practicality.

• •Proposes AnyMS, a training-free framework for multi-subject image customization.
• •Employs a bottom-up dual-level attention decoupling mechanism for text alignment, identity preservation, and layout control.
• •Utilizes pre-trained image adapters, eliminating the need for subject learning or adapter tuning.
• •Achieves state-of-the-art performance and supports complex compositions.