MagicFace: Slot-Driven High-Fidelity One-Shot Facial Appearance Editing

Facial Appearance Editing (FAE) focuses on modifying physical attributes like pose, expression, and lighting in facial images while preserving identity and background, which is crucial in photography. Despite significant progress, current research faces three main challenges: low generation fidelity, poor attribute preservation, and inefficient inference. To address these issues, this paper introduces MagicFace, a slot-driven, high-fidelity, one-shot FAE framework. Particularly, MagicFace employs Space-sensitive Physical Customization (SPC) for accurate query attribute transfer, utilizing rendering textures from the 3D Morphable Model (3DMM) to ensure fidelity and generalization. To preserve source attributes, we propose the Region-responsive Semantic Composition (RSC) module, guided by slot attention to learn and preserve decoupled source features. This approach can successfully maintain identity and reduces artifacts related to non-facial attributes such as hair, clothes, and background. Additionally, we introduce a consistency regularization technique to enhance editing controllability by leveraging prior knowledge from attention matrices in the diffusion model. Extensive experiments demonstrate that MagicFace achieves state-of-the-art performance in FAE task, outperforming existing methods.