As a detection paradigm proposed recently, DiffusionDet has exhibited its promising capability via formulating object detection as a denoising diffusion process, endowed with being trained once for all inferences. However, it suffers from slow convergence and valueless feature due to intuitive training style and redundant candidate boxes. To mitigate these issues, we proposed Guided DiffusionDet with novel Guided Diffusion Step and flexible Resample Mechanism. Technically, Guided Diffusion Step guides the decoder to denoise from candidate boxes with different levels of noise. This benefits to better refine the positions and sizes of these boxes compared to the naive training scheme of DiffusionDet. Besides, the proposed Resample Mechanism significantly eliminates those noisy boxes with little feature of target objects. In contrast with the DiffusionDet, our Guided DiffusionDet attains \({\textbf {1.4}}\) AP and \({\textbf {1.5}}\) AP gains respectively on COCO dataset and LVIS dataset.