Scholarlink

<p>Modeling long-range correlations and reconstructing fine details are essential for high-fidelity image inpainting. However, existing methods often face challenges of high computational overhead and over-smoothed outputs. This paper presents GCF-former, a novel framework that enhances inpainting performance via dynamic frequency-domain filtering. We introduce the Fused Adaptive Frequency Mixer (FAFM), which efficiently captures long-range dependencies and refines semantic features in Fourier space, reducing computational complexity to log-linear. We further propose a Multi-Scale Gated Convolution (MSGC), deployed in shallow layers to aggregate multi-scale features and enhance intricate detail recovery without incurring excessive parameter overhead. We evaluate GCF-former on the CelebA-HQ, Places2, and Dunhuang Challenge datasets. Results demonstrate that our framework outperforms other advanced methods in inpainting quality with lower computational cost.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Contextual aggregation and frequency filtering-enhanced long-range modeling for high-fidelity inpainting

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Wenyi Zhang,
  • Shengrong Zhao,
  • Jianhua Dong,
  • Hu Liang

摘要

Modeling long-range correlations and reconstructing fine details are essential for high-fidelity image inpainting. However, existing methods often face challenges of high computational overhead and over-smoothed outputs. This paper presents GCF-former, a novel framework that enhances inpainting performance via dynamic frequency-domain filtering. We introduce the Fused Adaptive Frequency Mixer (FAFM), which efficiently captures long-range dependencies and refines semantic features in Fourier space, reducing computational complexity to log-linear. We further propose a Multi-Scale Gated Convolution (MSGC), deployed in shallow layers to aggregate multi-scale features and enhance intricate detail recovery without incurring excessive parameter overhead. We evaluate GCF-former on the CelebA-HQ, Places2, and Dunhuang Challenge datasets. Results demonstrate that our framework outperforms other advanced methods in inpainting quality with lower computational cost.