Diffusion models have demonstrated impressive potential in semantic segmentation tasks. However, these models cannot accurately segment hidden polyps with complex structures owing to the absence of multi-scale conditional features for guiding the reverse process of diffusion models. To address this issue, we propose a dynamic multi-scale conditional diffusion model (PolypSegDiff). First, we design a dynamic multi-scale integration module to fuse the noise segmentation mask and the original image, dynamically extract multi-scale conditional features, and strengthen the network’s ability of identifying polyp areas. Second, we design a hierarchical feature enhancement module to extract and combine image features at different levels. This module significantly enriches the semantic diversity of conditional features, enabling the denoising network to more accurately understand the semantic relationships between polyps and the surrounding normal tissues. Experimental results across five publicly available polyp segmentation datasets demonstrate that PolypSegDiff outperforms existing popular methods in segmentation accuracy, achieving outstanding performance and robust generalization.