SPIDER: Boosting Blind Face Restoration via Simultaneous Prior Injection and Degradation Removal

Jie Wang, Yichen Yue, Dan Zeng

Keywords: Blind Face Restoration, Super-resolution, Degradation Removal, Diffusion Priors

Abstract: Existing blind face restoration (BFR) methods suffer from drastic performance drop under severe degradations. A common strategy is to first remove degradations and then restore the face by fully harnessing generative models. However, this sequential pipeline risks discarding subtle but crucial cues from already limited low-quality (LQ) inputs. To address this, we introduce a new learning paradigm: simultaneous prior injection and degradation removal (SPIDER). Unlike prior approaches, SPIDER injects semantic priors before degradation removal, thereby preserving identity-relevant features and mitigating the impact of corrupted LQ features. SPIDER consists of two key modules: (1) a prior injection module that distills purified degradation-unaware semantic control tokens from vision-language models, and (2) a degradation removal module equipped with an image-to-text degradation mapper and a degradation remover that refines distorted features into robust representations. This design leads to boosted BFR performance. Extensive experiments on both synthetic and real-world datasets, including challenging surveillance scenarios, demonstrate SPIDER's clear superiority over state-of-the-art BFR methods.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 24947