Existing training-free image composition methods based on diffusion models typically embed the feature information of foreground images into specified regions of the background image to guide the composition process. However, this embedding method often disrupts the denoising process of diffusion models, leading to inconsistency between foreground and background, as well as poor semantic alignment. To address these issues, we propose a novel training-free image composition method that incorporates two modules: complementary feature fusion embedding (CFFE) and adaptive feature reorganization (AFR). First, CFFE introduces composite image features extracted by the U-Net self-attention mechanism. These features are constructed from both foreground and background image information, enabling the preservation of foreground guidance while compensating for the loss of background semantic information in traditional embedding methods. Subsequently, the composite image and foreground fusion features are embedded to guide the composition process, while the number of embedded features is regulated to reduce potential composition biases. Furthermore, to address artifacts in the image transition regions caused by feature embedding, we introduce the AFR strategy. This strategy analyzes the covariance relationships of adjacent features to identify and replace abnormal features that cause discontinuous artifacts for enhancing image coherence. Experimental results demonstrate that the proposed method effectively improves foreground-background consistency and semantic alignment in synthesized images while achieving more harmonious image composition. This work also provides an effective solution for training-free image composition tasks.