Efficient LoFTR is a novel method for efficiently producing semi-dense matches across images. It improves upon the detector-free matcher LoFTR, which has shown strong matching capabilities in challenging scenarios but suffers from low efficiency. The paper introduces an aggregated attention mechanism for efficient feature transformation and a two-stage correlation layer for accurate subpixel correspondence refinement. The proposed method achieves significantly better efficiency than LoFTR, with a speed improvement of approximately 2.5 times, and can even surpass the state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue in terms of accuracy. The method is evaluated on multiple tasks, including homography estimation, relative pose recovery, and visual localization, demonstrating superior performance and efficiency. The key innovations include an aggregated attention mechanism that reduces redundant computations and a two-stage correlation layer that improves accuracy by refining matches at the subpixel level. The method is efficient and accurate, making it suitable for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction.Efficient LoFTR is a novel method for efficiently producing semi-dense matches across images. It improves upon the detector-free matcher LoFTR, which has shown strong matching capabilities in challenging scenarios but suffers from low efficiency. The paper introduces an aggregated attention mechanism for efficient feature transformation and a two-stage correlation layer for accurate subpixel correspondence refinement. The proposed method achieves significantly better efficiency than LoFTR, with a speed improvement of approximately 2.5 times, and can even surpass the state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue in terms of accuracy. The method is evaluated on multiple tasks, including homography estimation, relative pose recovery, and visual localization, demonstrating superior performance and efficiency. The key innovations include an aggregated attention mechanism that reduces redundant computations and a two-stage correlation layer that improves accuracy by refining matches at the subpixel level. The method is efficient and accurate, making it suitable for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction.