Some questions about training u-vit

[ximinng]

Dear Bao. Thanks for sharing your work! Some questions I'd like to ask after reading your code.

Since you have provided a series of model parameters (including autoencoder_kl.pth, caption_decoder.pth and image ViT-B/32 CLIP encoder, etc), only the U-Vit model needs to be trained for downstream task application, am I understanding correctly? So, can I train my u-vit model via train_t2i_discrete.py in U-Vit code and integrate it in unidiffuser directly?

Much appreciate if you can take a look sometime. :)

[baofff]

yes

[ximinng]

i have a question about classifier free guides in unidffuser, what is the null_context in of cfg in unidffuser, text prompts or img, since unidffuser is able to execute t2i and i2t ?