HifiFace: 3D Shape and Semantic Prior Guided High Fidelity Face Swapping

Facial area swapping is the endeavor of building visuals with resource face identification and the characteristics from the concentrate on image. Existing techniques face problems in providing photo-practical outcomes or conserving the initial face shape.

Facial area recognition – creative interpretation in Hollywood CA. Impression credit history: YO! What Occurred To Peace? via Flickr, CC BY-SA 2.

A current research proposes a novel conclude-to-conclude mastering framework, which can protect the face shape and deliver high fidelity face-swapping outcomes. A 3D shape-aware identification extractor generates an identification vector with specific shape facts to implement precise face shape transfer. A semantic facial fusion module is proposed to reach a far better blend in feature-degree and image-degree. That can help to fix occlusion and lights challenges.

Considerable experiments exhibit that the instructed approach can deliver increased fidelity outcomes than previous techniques. It can be applied in the film market, computer game titles, or face forgery detection.

In this operate, we propose a high fidelity face swapping approach, identified as HifiFace, which can very well protect the face shape of the resource face and deliver photo-practical outcomes. Not like other current face swapping performs that only use face recognition model to retain the identification similarity, we propose 3D shape-aware identification to handle the face shape with the geometric supervision from 3DMM and 3D face reconstruction approach. In the meantime, we introduce the Semantic Facial Fusion module to enhance the blend of encoder and decoder capabilities and make adaptive mixing, which can make the outcomes more photo-practical. Considerable experiments on faces in the wild exhibit that our approach can protect far better identification, primarily on the face shape, and can deliver more photo-practical outcomes than previous condition-of-the-artwork methods.

Analysis paper: Wang, Y., “HifiFace: 3D Form and Semantic Prior Guided Significant Fidelity Facial area Swapping”, 2021 . Hyperlink to the posting: https://arxiv.org/abdominal muscles/2106.09965