Skip to content

Masked Diffusion Transformer (MDT)

Predict masked tokens for image synthesis

Visit official site

What MDT is

The model is described as a masked diffusion transformer that improves how diffusion probabilistic models learn relations between object parts in images.

What you can do with it

  • Generate high resolution images
  • Work on fast learning in image synthesis
  • Use MDTv2 to improve FID scores in image synthesis

Model notes

  • Training masks some tokens in latent space
  • An asymmetric transformer predicts masked tokens
  • The diffusion generation process is kept

Pros & cons

✓ What we like

  • The diffusion generation process is kept during training
  • Token masking happens in latent space

! What to watch out for

  • Prediction relies on the unmasked tokens
  • MDTv2 changes the macro network structure and training strategy

FAQ

What does it mask?

The material states some tokens are masked in latent space during training.

What does MDTv2 change?

A more efficient macro network structure and training strategy are named.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review