ENTED: Enhanced Neural Texture Extraction and Distribution for Reference-based Blind Face Restoration

We present ENTED, a new framework for blind face restoration that aims torestore high-quality and realistic portrait images. Our method involvesrepairing a single degraded input image using a high-quality reference image.We utilize a texture extraction and distribution framework to transferhigh-quality texture features between the degraded input and reference image.However, the StyleGAN-like architecture in our framework requires high-qualitylatent codes to generate realistic images. The latent code extracted from thedegraded input image often contains corrupted features, making it difficult toalign the semantic information from the input with the high-quality texturesfrom the reference. To overcome this challenge, we employ two specialtechniques. The first technique, inspired by vector quantization, replacescorrupted semantic features with high-quality code words. The second techniquegenerates style codes that carry photorealistic texture information from a moreinformative latent space developed using the high-quality features in thereference image's manifold. Extensive experiments conducted on synthetic andreal-world datasets demonstrate that our method produces results with morerealistic contextual details and outperforms state-of-the-art methods. Athorough ablation study confirms the effectiveness of each proposed module.