The Devil Is in the Leakage: A Disentangled Dual-Purification Framework for High-Fidelity Hairstyle Transfer

Hairstyle transfer aims to transplant a reference hairstyle onto a source subject while maintaining the source's identity. Existing diffusion-based methods often struggle with 'leakage' problems: identity information from the reference hair accidentally transfers to the face, or residual geometric artifacts from the 'bald' source image distort the final hair contour. This paper investigates how to systematically disentangle these features to achieve high-fidelity results.

The authors propose the Dual-Purification Framework (DPF), which introduces two specialized training-time regularizers to clean the information streams:

1. Adversarial Hairstyle Purification (AHP): This module uses an adversarial game between a generator and a discriminator to minimize the mutual information between the hairstyle features and the source identity. By forcing the discriminator into a state of maximal uncertainty, the model effectively strips identity-related cues from the hairstyle representation.
2. Contrastive Geometric Purification (CGP): This module addresses 'Flaw Leakage in Bald' by using a contrastive objective. It forces the ControlNet to produce divergent geometric features for different hairstyles, preventing the model from relying on subtle artifacts (like shadows or hair remnants) in the bald source image as a shortcut for generation.

These modules are applied only during training, ensuring that the inference process remains efficient without adding computational overhead.

By formalizing the leakage problem into two distinct components—Identity Leakage and Flaw Leakage—the authors provide a clear path to improving local editing in foundation models. The DPF approach demonstrates that rather than just scaling up models, explicit purification of latent streams is necessary to achieve the precise disentanglement required for high-fidelity image manipulation. This framework sets a new state-of-the-art for hairstyle transfer, enabling more robust results across diverse poses and subjects.