基于 APPLE 的属性保持伪标签人脸换脸
摘要
人脸换脸旨在将 source face 的 identity 转移到 target face 上,同时保持 target-specific attributes,如 pose、expression、lighting、skin tone 和 makeup。然而,由于无法获得 face swapping 的 real ground truth,实现 accurate identity transfer 和 high-quality attribute preservation 仍然具有挑战性。最近的基于 diffusion 的方法通过对 masked target images 进行 conditional inpainting 来提高视觉保真度,但 masked condition 会移除 crucial appearance cues,导致生成 plausible 但 misaligned attributes。为解决这一限制,我们提出 APPLE (Attribute-Preserving Pseudo-Labeling),一种 fully diffusion-based 的 teacher-student 框架,用于 attribute-preserving face swapping。我们的做法引入 teacher design,通过 (1) 条件 deblurring 公式提高 global attributes 如 skin tone 和 illumination 的保持,以及 (2) attribute-aware inversion scheme 进一步增强 fine-grained attribute preservation 如 makeup。APPLE 将 student 条件于 clean pseudo-labels 而非受损 masked inputs,从而实现更可信的 attribute preservation。结果表明,APPLE 在 attribute preservation 上实现 state-of-the-art performance,同时保持 competitive identity transferability。
引用
@article{arxiv.2601.15288,
title = {APPLE: Attribute-Preserving Pseudo-Labeling for Diffusion-Based Face Swapping},
author = {Jiwon Kang and Yeji Choi and JoungBin Lee and Wooseok Jang and Jinhyeok Choi and Taekeun Kang and Yongjae Park and Myungin Kim and Seungryong Kim},
journal= {arXiv preprint arXiv:2601.15288},
year = {2026}
}
备注
Accepted at CVPR 2026. Project Page: https://cvlab-kaist.github.io/APPLE/