Tri-Modal Masked Diffusion Models Pretrained from Scratch
October 13, 2026
This research introduces Masked Diffusion Models (MDM) pretrained simultaneously on text, image-text, and audio-text data. The study provides analysis of multimodal scaling laws, modality mixing ratios, and optimized SDE-based inference sampling.
HOW THIS AFFECTS YOU
●
builderYou can use the optimized inference sampling defaults to improve multimodal generation performance.
●
researcherYou can apply these findings on modality mixing and noise schedules to improve multimodal model training.