Missing MRI modalities is a frequent challenge in clinical brain tumor imaging, limiting the effectiveness of multimodal segmentation models. In this work, we propose an ensemble framework for synthesizing missing MRI modalities from the available ones. It combines a Modality Translation Encoder-Decoder and a Modality Translation Brownian Bridge Diffusion Model, both operating in a compact latent space generated by a pretrained Volumetric Compression Network. This design enables whole-volume 3D synthesis with moderate computational demands and improved anatomical coherence. Each model is trained and validated on the BraSyn 2025 dataset, and their outputs are fused to increase robustness against structural and contrast variability. Evaluation on the validation set shows that the synthetic images produced by our ensemble closely resemble the original missing modalities and, when combined with the available ones, support effective tumor segmentation, demonstrating the method’s effectiveness for clinical data completion.