DreamID: High-Fidelity and Fast diffusion-based Face Swapping via Triplet ID Group Learning

In this paper, we introduce DreamID, a diffusion-based face swapping modelthat achieves high levels of ID similarity, attribute preservation, imagefidelity, and fast inference speed. Unlike the typical face swapping trainingprocess, which often relies on implicit supervision and struggles to achievesatisfactory results. DreamID establishes explicit supervision for faceswapping by constructing Triplet ID Group data, significantly enhancingidentity similarity and attribute preservation. The iterative nature ofdiffusion models poses challenges for utilizing efficient image-space lossfunctions, as performing time-consuming multi-step sampling to obtain thegenerated image during training is impractical. To address this issue, weleverage the accelerated diffusion model SD Turbo, reducing the inference stepsto a single iteration, enabling efficient pixel-level end-to-end training withexplicit Triplet ID Group supervision. Additionally, we propose an improveddiffusion-based model architecture comprising SwapNet, FaceNet, and ID Adapter.This robust architecture fully unlocks the power of the Triplet ID Groupexplicit supervision. Finally, to further extend our method, we explicitlymodify the Triplet ID Group data during training to fine-tune and preservespecific attributes, such as glasses and face shape. Extensive experimentsdemonstrate that DreamID outperforms state-of-the-art methods in terms ofidentity similarity, pose and expression preservation, and image fidelity.Overall, DreamID achieves high-quality face swapping results at 512*512resolution in just 0.6 seconds and performs exceptionally well in challengingscenarios such as complex lighting, large angles, and occlusions.