Tri-Modal Masked Diffusion Models: SDE Reparameterization Decouples Batch Size for Unified Text-Image-Audio Generation at 3B Scale

Researchers pretrained the first tri-modal Masked Diffusion Model (MDM) from scratch on text, image-text, and audio-text data, achieving strong generation capabilities at a 3B parameter scale. The study introduces an SDE-based reparameterization that decouples physical batch size from logical gradient variance, eliminating the need for manual batch tuning. This work represents the largest systematic open study of multimodal discrete diffusion models to date, offering critical insights into scaling laws across modalities.