Denoising diffusion probabilistic models (DDPMs) have shown promise for generating high-resolution synthetic images. In medical imaging, there is a growing demand for both realistic image synthesis and deterministic outcomes that can guide downstream applications effectively. In this study, we propose MED-INPAINT, an adaptable multi-level conditional DDPM framework. MED-INPAINT incorporates contrast priors for accelerated sampling and performs inpainting of pelvic magnetic resonance imaging (MRI) scans, enabling high-quality image synthesis with reasonably low uncertainty. Our results highlight the effectiveness of MED-INPAINT in generating realistic and detailed pelvic MRI images, assessing its uncertainty using various denoising steps at inference. MED-INPAINT outperformed baseline U-Net and cycle-consistent generative adversarial network (Cycle-GAN) models, demonstrating its potential for various medical imaging applications.