NLKFill: High-Resolution Image Inpainting With a Novel Large Kernel Attention

The paper introduces a novel single-stage network called NLKFill, which uses large kernel attention (LKA) to improve high-resolution image inpainting. This approach effectively combines the strengths of convolutional neural networks (CNNs) and transformers to capture both global and local image details, overcoming the limitations of transformers in handling high-resolution images. NLKFill reduces parameters, enhances inference speed, and allows direct training on 1024 × 1024 resolution images. It demonstrates superior performance in extracting global high-frequency and local details, showing excellent generalization across various datasets, including Places2, Celeba-HQ, FFHQ, and NVIDIA's random irregular mask dataset Pconv.