An Efficient Diffusion for Blind Face Restoration
Yu Wang, Gencheng Wang, Rong Chen, Yuzhen Chen · 2024
In recent years, many deep learning-based methods especially the diffusion model have made a remarkable breakthrough in blind face restoration. Thanks to the powerful generative capabilities of the diffusion model, these methods can generate relatively better results when confronted with intricate degradation without multiple constraints. However, the inference efficiency of diffusion models is hampered by the iterative sampling process inherent in model inference. To address this problem, we focus on the lightweight and fast diffusion model. In this study, we propose an efficient diffusion for blind face restoration (EDBFR). The key of our method is to introduce a specialized diffusion model solver which crafted to refine the diffusion partial differential equations to enhance the sampling speed. Additionally, we devise a lightweight residual module and integrate it into the Unet network of diffusion models to reduce the number of parameters while maintaining performance. Comprehensive experiments on CelebA-Test, LFW-Test and WebPhoto-Test demonstrate that our proposed method surpasses the state-of-the-art deep learning-based blind face restoration methods both quantitatively and qualitatively. Especially, compared with the state-of-the-art diffusion models, our model has 20% inference time, and 50% parameters.