Do you have one SeeCoder that provides alignment with T5 rather than CLIP?
fiona-lxd opened this issue · comments
fiona-lxd commented
Do you have one SeeCoder that provides alignment with T5 rather than CLIP?
Xingqian Xu commented
So far no, but a SeeCoder can be obtained following the same training we mentioned in the paper by just changing the underlying T2I model into a T5-supported T2I model.
fiona-lxd commented
I see. Thank you for the quick reply. BTW, how much time does it cost to train the seecoder with 16 A100?
bonlime commented
@xingqian2018 Could you please share the training time on 16 A100? I'm thinking of replicating/finetuning the model and want some estimates of how long it would take