Facial landmark alignment is a key prerequisite for most existing hair attribute editing algorithms. However, when this alignment fails, especially in privacy-preserving scenarios, it introduces critical technical bottlenecks that hinder practical applications. Traditional approaches heavily rely on facial keypoint detection to establish geometric constraints. When facial information is occluded, such as through mosaic obfuscation (as shown in Fig. 1), mainstream methods suffer from landmark detection failures and inaccurate contour estimations, leading to structural dislocation and texture distortion in hair editing results. To address this critical challenge, we propose CoifNet, a novel framework designed for mosaic-obscured faces. The contributions of CoifNet are twofold: (i) A Facial Outline Fitting (FLF) module that leverages an average face model to learn a mapping from blurred to clear facial contours. This enables accurate facial geometry estimation even under zero-keypoint conditions, effectively mitigating system failures due to landmark loss; (ii) A novel texture encoder, TexPro, which incorporates a texture modulation unit to disentangle and enhance texture representation in latent space, thereby improving the modeling of hair texture. Specifically, the FLF module aligns the standard face model with the mosaic region to generate adaptive facial contours that guide the transfer of hairstyle between the input and target images. TexPro innovatively employs high-frequency auxiliary inputs to modulate texture encoding, enabling the generation of more diverse and realistic texture patterns. Extensive experiments on the K-Hairstyle and CelebAMask-HQ datasets demonstrate the effectiveness of our approach. Notably, CoifNet reduces the FID score on the mosaic K-Hairstyle dataset from 26.84 to 18.06, producing results with richer detail and higher realism.
