Towards consistent and controllable image synthesis for identity-preserving face editing

Face editing involves modifying facial attributes like expression, head pose, or lighting, with the goal of preserving the subject’s unique identity features. Diffusion models have recently emerged as the dominant approach in visual generation, driven by their strong generative power. However, challenges persist in the realm of face editing, where independently and correctly editing target attributes while preserving high-fidelity identity information remains a formidable problem. In this paper, we present RigFace, a novel framework that combines controllable signals derived from a 3D Morphable Model (3DMM) with a fine-tuned Stable Diffusion (SD) model. Our basic idea to achieve by leveraging disentangled facial attributes provided by 3DMM and harnessing the strong generative capacity of Stable Diffusion. Specifically, our method contains: 1) A Spatial Attribute Encoder that provides robust and decoupled conditions of background, pose, expression and lighting; 2) A FaceFusion module that transfers identity information at different resolutions from the Identity Encoder to the Denoising UNet of a pre-trained SD model through self-attention, which facilitates detailed identity preservation. Our model achieves superior performance in both identity preservation and photorealism compared to existing face editing models. © 2026 The Authors