Enhanced conditional diffusion network for facial animation generation

With the increasing demand for realistic facial animation in applications such as virtual reality, film, and video conferencing, existing methods face challenges in preserving high-fidelity details and accurate motion dynamics. Traditional GAN-based approaches often introduce artifacts and lack the precision required for natural facial expressions. To address these challenges, we propose an Enhanced Conditional Diffusion Network (ECDNet), which integrates a diffusion-based denoising model with Efficient Channel Attention (ECA) and Coordinate Attention (CA) mechanisms within an Adaptive Graph Convolutional Network (AGCN). This architecture enables ECDNet to effectively capture complex facial features and accurately transfer appearance and motion conditions, thereby improving the quality of generated animations. Additionally, ECDNet employs deepwise separable convolutions, which enhance computational efficiency while preserving fine-grained facial details, further improving the quality of the generated animations. Experimental results on the VoxCeleb, VoxCeleb2, and CelebA datasets demonstrate that ECDNet achieves significant improvements, confirming its effectiveness in generating high-quality, realistic facial animations.