Predict-and-refine conditional diffusion model for blind face restoration

doi:10.1117/1.JEI.34.6.063010

Predict-and-refine conditional diffusion model for blind face restoration

2025-11-04

0

PRE

AI

DOI：10.1117/1.JEI.34.6.063010

原文链接

原文求助

摘要

En 中文

Diffusion models have demonstrated powerful performance in blind face restoration (BFR). However, they primarily encounter two limitations. First, their performance typically declines when encountering intricate degradation beyond the scope of training datasets. Second, these methodologies rely on multiple constraints (e.g., fidelity, perceptual quality, and adversarial losses) to guide the improvement of model performance. In this paper, we propose a predict-and-refine conditional diffusion model, which provides a relatively reliable principle to guide the model design and yields a new insight into the interpretable diffusion for BFR. Specifically, we incorporate the existing pre-trained restoration model into a diffusion model based on mathematical assumptions and derivation. We utilize a pre-trained restoration model as an initial predictor from the natural image distribution while maintaining the input image content. Then, the initial predictor is incorporated into the diffusion model to guide the reverse denoising process, which bridges the gap between the generative capability of the diffusion priors and the effectiveness of the initial predictor. Extensive experiments on both simulation datasets and real datasets demonstrate that the proposed method exhibits excellent performance in terms of both visual effects and objective evaluation metrics. (c) 2025 SPIE and IS&T

Keyword:

diffusion model

blind face restoration

pre-trained

gradient loss

Predict-and-refine conditional diffusion model for blind face restoration

摘要

期刊

机构

引用论文