kayariyan28/ghost-score-face-swap

Ghost-Score-Driven Multi-Pipeline Face Swap is a local research prototype written in Python that replaces one person's face in a photo with another person's face. Rather than using a single approach, it routes each swap through one of three pipelines depending on how well the transfer went, using a metric called the ghost score to decide. The ghost score is a simple measurement: it compares how similar the output face looks to the source identity versus the target identity. A negative score means the swap succeeded, the result looks more like the source. A positive score signals leakage, meaning the original target face is still showing through, and the system escalates to a stronger recovery pipeline. The three pipelines are Classical Compositing, AI Synthesis, and Pro Adaptive. Classical Compositing focuses on preserving fine pixel detail and texture from the source. AI Synthesis is faster and more tolerant of difficult head poses. Pro Adaptive is the most powerful: it runs multiple candidate approaches and applies a 3D-aware visible-surface mask to avoid blending parts of the face that are hidden by occlusion, such as hair or other objects in front of the face. The tool is designed for local execution only, there is no cloud API and no telemetry. A 200-pair benchmark reported in the accompanying research paper found that the Pro Adaptive mode reached a 92.5 percent identity match rate, outperforming the simpler pipelines. The research paper is published on Zenodo (DOI 10.5281/zenodo.20179682). The README states the runtime is tuned for macOS on Apple Silicon, though Linux can run the tool as well.