One block-causal MMDiT that models images and text in the same latent sequence, in both directions. The image branch reuses this repo's VAE; text gets its own VAE, so a caption becomes a sequence of ...
Some results have been hidden because they may be inaccessible to you
