CrossSwap: A shape-aware identity cross-attention face swapping framework
Hojun Seo, Junyeoung Ahn, Byungseok Kang · 2026
Read CrossSwap: A shape-aware identity cross-attention face swapping framework by Hojun Seo, Junyeoung Ahn, Byungseok Kang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0349356.txt.
Description
Hojun Seo, Junyeoung Ahn, and Byungseok Kang present CrossSwap, a face swapping system designed to fix the uncanny results of earlier tools, which often lost the original person's identity. Their model studies the source face's shape before blending, then uses attention blocks that adjust across scales to pour identity features into the target picture while keeping the target's pose and expression.
Against three recent rival methods, CrossSwap took first place on pose and expression accuracy and second on identity retention and overall image quality. The authors argue that respecting face shape, not just surface features, is what makes a swap look like the same person in a new scene. The work sits inside generative image research and carries the field's usual dual-use questions.
Catalog record
| CREATOR | Hojun Seo, Junyeoung Ahn, Byungseok Kang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45129 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |