VolumeSwap: Volumetric Decomposed 3D‐Aware Face Swapping

Abstract 3D face swapping has been widely applied in entertainment, gaming and privacy protection. Traditional methods often interpolate directly in the latent space to generate a global swapped code, followed by per‐image latent code optimisation. While these approaches achieve 3D consistency, they frequently struggle with preserving the source identity and target attributes and are burdened by unstable and extensive optimisation processes. In our work, we adopt a novel volumetric perspective, deconstructing the 3D‐aware face swapping problem into two distinct volumes: the inner swapped face and the background. We introduce VolumeSwap, featuring dual branches specifically tailored for each volume. The face module integrates explicit semantic and geometric guidance into the face swapping process, significantly enhancing attribute transfer. By separately modelling the face and background, we accelerate the optimisation process, minimising the background's undesirable impact on the facial region and resulting in clearer, more precise swapped faces. Our method not only reduces optimisation time to approximately 1 min (compared to 5 min of SOTA) but also ensures high fidelity in both the inner face and background reconstruction. Extensive experimental evidence demonstrates the superiority of our approach over existing state‐of‐the‐art methods, both in terms of swapping quality and efficiency.