full and remaining unmasked latent embeddings

sail-sg / MDT

Masked Diffusion Transformer is the SOTA for image synthesis. (ICCV 2023)

Apache License 2.0

488 stars 36 forks source link

full and remaining unmasked latent embeddings #37

Open ZhaoLiruii opened 3 months ago

ZhaoLiruii commented 3 months ago

Hello, I would like to know how this sentence in the paper is achieved: "In MDT, both the full and remaining unmasked latent embeddings are fed into the diffusion model during training."? I can't find this part in the code

ZhaoLiruii commented 3 months ago

@gasvn

gasvn commented 3 months ago

https://github.com/sail-sg/MDT/blob/4b8bcbb656750cb8b763c309711234652b95b1c5/masked_diffusion/train_util.py#L227 Check this function.