LaCoSwap: One-Step High-Fidelity Face Swapping via Latent Consistency Model

Recently, diffusion models have demonstrated exceptional fidelity and control for face-swapping. However, existing approaches suffer from limited inference speed due to the requirement for a large number of iterative steps. In this paper, we solve this issue by proposing LaCoSwap, the first efficient and effective latent consistency model-based face-swapping method, enabling one-step face-swapping with high-quality results. Specifically, we introduce a face-swapping latent consistency distillation process, tailored for face-swapping by optimizing classifier-free guidance and incorporating identity constraints to generate high-fidelity swapped faces. The consistency constraint is employed to distill a consistency model from a well-designed teacher model based on diffusion models, resulting in superior performance in the distilled LaCoSwap. During the sampling process, to further elevate the quality of the swapped face, we propose a multistep masked consistency sampling strategy, allowing full control and customization of the mask and conditions. Our experiments demonstrate that LaCoSwap achieves over 200 times faster inference speeds than the traditional diffusion-based methods. The results highlight LaCoSwap's exceptional performance in generating natural and high-quality faces that better preserve identity and attributes, surpassing benchmarks established by traditional methods.