sail-sg / MDT

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

will you release a pretrained checkpoint of B/2 model? #45

Open aaab8b opened 3 months ago

aaab8b commented 3 months ago

Thanks for the wonderful work again and I still wants to do some experiments on the small models like B/2 or S/2 due to my limited resources. Will you guys release a pretrained checkpoint of B/2 model or S/2 model?

gasvn commented 3 months ago

Sorry I no longer have these checkpoints since I no longer have the access to the clusters used to train our model.

aaab8b commented 3 months ago

Sorry I no longer have these checkpoints since I no longer have the access to the clusters used to train our model.

I use the same 32x8 batchsize and check every single detail and cannot get the same fid result for training my s/2 and b/2 models. The performance drops about ~20 fid score on my s/2 model. And the only thing I'm not certain about is your dynamic ratio. the code its from 0.3 to 0.5. but in your paper it's 0.3 to 0.05. which is correct?

gasvn commented 3 months ago

The mask ratio is from 0.3 to 0.5. Did you check if you are using the cfg for testing?