Magnetic resonance imaging (MRI) provides spatial, multi-channel volumetric data, where each channel captures complementary features of underlying anatomy or pathology. However, in practice, one or more of these channels can be missing due to time constraints and motion artifacts. This limits the applicability of current MRI-based diagnosis pipelines and motivates the need for robust MRI synthesis. In this work, we trained a unified diffusion framework (DiffuseMRI) for controllable MRI sequence synthesis based on DiffuseTME, designed to handle high-dimensional, spatially aligned volumetric data with complex inter-channel relationships. Our model features two key components: (1) a hierarchical feature injection mechanism that enables multi-resolution conditioning on spatially aligned observed MRI sequences, and (2) channel-wise attention modules to model the dependencies across MRI sequences. To generalize across arbitrary missing sequences, we train the model using a random masking strategy, allowing DiffuseMRI to reconstruct any missing sequences while preserving anatomical consistency. Our framework is evaluated on the BraSyn benchmark, part of BraTS challenge in MICCAI 2025, which standardizes the evaluation of MRI synthesis methods for downstream brain tumor segmentation.