Text guided image diffusion model has demonstrated remarkable ability in consistent image generation. In this paper, we introduce a training free image composition framework that realizes the non-rigid objects composition based on a pair of source and target prompts. Specifically, we aim at blending the user provided object reference image into the background image in a non-rigid manner and keep the balance of fidelity and editability. For example, we can make a standing dog jumping while preserving its shape and appearance under the guidance of target prompt. Our proposed method has three key components: firstly, the reference image and background are inverted into latent noises with different image inversion methods. Secondly, we guarantee the consistent image attribute generation of the reference object by injecting the self-attention key and value features from original pipeline in sampling steps. Thirdly, we iteratively optimize the object mask in the target pipeline, and progressively compose image in different regions. Experiments shows that our proposed method can achieve the non-rigid object image editing and seamless composition, the results are impressive in consistent and editable image composition.