Reproduce “VQFR: Blind Face Restoration with Vector-Quantized Dictionary and Parallel Decoder” on OpenResearch

About this paper

Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, Ming-Ming Cheng

VQFR introduces a novel blind face restoration framework utilizing a vector-quantized (VQ) codebook for generating high-quality facial details and a parallel decoder architecture to preserve fidelity. The method achieves superior perceptual quality with improved fine-grained details while maintaining facial identity and expression, outperforming state-of-the-art techniques on synthetic and real-world datasets.

Key insights

• A VQ codebook, when trained on high-quality data, can implicitly remove degradations from low-quality inputs because its codes lack degradation-related information.
• The compression patch size (`f`) of the VQ codebook is critical; `f=32` for 512x512 images allows for high-quality detail generation but necessitates architectural solutions for fidelity preservation.
• Decoupling the generation of realistic textures from the preservation of input fidelity via a parallel decoder effectively mitigates the conflict between quality and identity/expression retention.

Summary by alphaXiv.