search

sail-sg / MDT

Masked Diffusion Transformer is the SOTA for image synthesis. (ICCV 2023)
Apache License 2.0
500 stars 35 forks source link

MDT-G ? #3

Closed achen46 closed 1 year ago

achen46 commented 1 year ago

Thanks for great work. In the paper, best FID results are for MDT-G. But I cannot find the config in here.

Can you let us know the exact config for MDT-G ?

gasvn commented 1 year ago

-G denotes the sampling with classifier-free guidance. It's not a new model but a sampling scheme.

achen46 commented 1 year ago

@gasvn Thanks for the response. May I ask what is the classifier-free guidance value you used ?

achen46 commented 1 year ago

@gasvn Also I wonder, why you choose to use a total batch size of 256 ? it seems to be more efficient to use a larger batch size like 512 but are there any reasons behind this decision ?

gasvn commented 1 year ago

@gasvn Thanks for the response. May I ask what is the classifier-free guidance value you used ?

The sampling parameters are in https://github.com/sail-sg/MDT/blob/main/scripts/image_sample.py

You can use the script in https://github.com/sail-sg/MDT#evaluation to evalute the model.

gasvn commented 1 year ago

@gasvn Also I wonder, why you choose to use a total batch size of 256 ? it seems to be more efficient to use a larger batch size like 512 but are there any reasons behind this decision ?

Your are right. But to make fair comparisons with DiT, we also choose the batch size of 256.

achen46 commented 1 year ago

@gasvn Thanks for all the prompt answers.