Reproduce “A2BFR: Attribute-Aware Blind Face Restoration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenxin Zhu, Yushun Fang, Lu Liu, Shibo Yin, Xiaohong Liu, Xiaoyun Zhang, Qiang Hu, Guangtao Zhai
The A²BFR framework unifies high-fidelity blind face restoration with fine-grained, prompt-controllable attribute generation in a single stage, addressing the ill-posed nature of face restoration and the lack of controllability in existing diffusion models. It achieves robust perceptual quality and precise attribute alignment by integrating image-text conditioning and a novel training strategy, outperforming current state-of-the-art methods in both fidelity and attribute accuracy.
Key insights
- Jointly conditioning a DiT-based model on both degraded visual inputs and explicit textual prompts through cross-modal attention effectively guides the denoising process for attribute-aware restoration.
- A supervised learning approach that explicitly aligns restored latent representations with ground-truth semantic attributes (AAL) and encourages attribute separation with a dual-training mechanism (SDT) is critical for high prompt adherence and fidelity.
- The construction of a large-scale, high-fidelity dataset (AttrFace-90K) with diverse, fine-grained attribute-paired images and captions provides essential supervision for controllable and semantically aligned face restoration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.