Core Concepts
Leveraging pretrained Stable Diffusion for blind face restoration achieves state-of-the-art performance.
Abstract
Blind face restoration is a crucial task in computer vision, addressing various degradations in face images. The proposed BFRffusion method utilizes the pretrained Stable Diffusion to extract features from low-quality images and restore realistic facial details. A privacy-preserving face dataset called PFHQ is introduced for training networks, addressing privacy and bias concerns. Extensive experiments demonstrate the superior performance of BFRffusion on synthetic and real-world datasets compared to existing methods. The architecture consists of modules like SDRM, MFEM, TTPM, and PDUM to guide the restoration process effectively.
Stats
Our BFRffusion achieves PSNR: 24.83 and SSIM: 0.7143 on CelebA-Test.
PFHQ dataset includes 60K paired face images for training blind face restoration networks.
Quotes
"Our BFRffusion achieves state-of-the-art performance on both synthetic and real-world public testing datasets."
"Our PFHQ dataset is an available resource for training blind face restoration networks."