Face blending is a fundamental post-processing step in virtually all modern face manipulation pipelines. After synthesizing or transforming a target face region (via deepfake synthesis, face2face reenactment, or faceswap), the result is blended back into the original image using techniques like alpha blending, Poisson blending, or learned blending methods. The blending boundary often leaves intrinsic image discrepancies that forensic methods can exploit.
Key papers
- 2019-li-face-xray — Proposes detecting face X-ray signatures of blending boundaries, exploiting the fact that blended images exhibit inconsistencies in image acquisition characteristics (sensor noise, compression artifacts) across the boundary; demonstrates that this universal blending property enables generalization across all manipulation methods.
Related topics
- face-manipulation (broader synthesis techniques)
- face-forgery-detection (detection of blended forgeries)
- image-blending (general image composition)