VQFR: Blind Face Restoration with Vector-Quantized Dictionary and Parallel Decoder

Although generative facial prior and geometric have recently demonstrated high-quality results for blind face restoration, producing fine-grained details faithful to inputs remains a challenging problem. Motivated by the classical dictionary-based methods recent vector quantization (VQ) technique, we propose VQ-based restoration method – VQFR. VQFR takes advantage of low-level feature banks extracted from faces can thus help recover realistic details. However, simple application VQ codebook cannot achieve good with identity preservation. Therefore, further introduce two special network designs. 1). We first investigate compression patch size in find that designed proper is crucial balance quality fidelity. 2). To fuse features while not “contaminating” generated codebook, proposed parallel decoder consisting texture main decoder. Those decoders then interact warping module deformable convolution. Equipped as detail dictionary design, largely enhance restored keeping fidelity previous methods.