From the source
Apple researchers introduced the first tri-modal Masked Diffusion Model (MDM) pretrained from scratch on text, image-text, and audio-text data.
They systematically analyzed multimodal scaling laws, modality mixing ratios, noise schedules, and batch-size effects, proposing a novel SDE-based reparameterization that decouples physical and logical batch sizes.
A preliminary 3B-parameter model trained on 6.4T tokens achieved strong results in text generation, text-to-image, and text-to-speech tasks.