Research map / Image generation / Editing and style transfer
BlendFace: Re-designing Identity Encoders for Face-Swapping
Kaede Shiohara, Xingchao Yang, Takafumi Taketomi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ArcFace with BlendFace for both source encoding and identity loss in the face-swapping model (ablation) | AEI-Net face-swapping model · FaceForensics++ | Gaze distance: 0.1450 -> 0.1301 | Improved |
| replace ArcFace with BlendFace for both source encoding and identity loss (ablation) | AEI-Net face-swapping model · FaceForensics++ | Pose distance: 0.0139 -> 0.0125 | Improved |
| replace ArcFace with BlendFace for both source encoding and identity loss (ablation) | AEI-Net face-swapping model · FaceForensics++ | ArcFace identity distance: 0.3314 -> 0.4630 | Hurt |
| use BlendFace only as source encoder (ArcFace loss) (ablation) | AEI-Net face-swapping model · FaceForensics++ | Gaze distance: 0.1450 -> 0.1376 | Improved |
| use BlendFace only as identity loss (ArcFace source) (ablation) | AEI-Net face-swapping model · FaceForensics++ | Gaze distance: 0.1450 -> 0.1377 | Improved |
| pretrain ArcFace with swapped faces at attribute replacement probability p of 0.25 or higher (ablation) | BlendFace encoder · VGGFace2 similarity distributions | similarity gap between Same and Swapped | Improved |
| train BlendFace with swapped faces versus plain ArcFace | BlendFace encoder · AgeDB verification | accuracy: 0.9838 -> 0.9813 | Hurt |
| use intersection mask instead of single mask when blending in pretraining (ablation) | BlendFace encoder · face-swapping qualitative | consistency of generated images | Improved |
| mask the source image at inference or training instead of using BlendFace (ablation) | Arc-Arc face-swapping model · FaceForensics++ | attribute consistency | No clear effect |
| BlendFace versus previous face-swapping models on shape similarity | BlendFace face-swapping model · FaceForensics++ | Shape distance: 0.2962 (InfoSwap) vs 0.3542 | Hurt |
Closest papers by meaning
- Learning Disentangled Representation for One-Shot Progressive Face Swapping (TPAMI 2024 · similarity 0.91)
- Reinforced Disentanglement for Face Swapping without Skip Connection (ICCV 2023 · similarity 0.90)
- Smooth-Swap: A Simple Enhancement for Face-Swapping With Smoothness (CVPR 2022 · similarity 0.90)
- Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion (CVPR 2025 · similarity 0.87)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.87)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
