semantic-consistency

LatentUMM introduces dual latent alignment to improve cross-modal consistency in unified multimodal models by aligning transformations and stabilizing latent dynamics.