DiT results on CIFAR10

[yuanzhi-zhu]

have you tried to run DiT on CIFAR10 dataset?
I did some simple expr and found that DiT does not work well on CIFAR10.

[tanghengjian]

i also found sample.py script always give same result image under same label.
In the workflow of DiTBlock, i wonder there is no cross attention , so i guess the variation ability may be a challange to DiT?

[yuanzhi-zhu]

i also found sample.py script always give same result image under same label.
In the workflow of DiTBlock, i wonder there is no cross attention , so i guess the variation ability may be a challange to DiT?

Hi @tanghengjian, I do not know if your question is related to the cifar expr, but did you change the seed in the sample.py script?

DiT/sample.py

Line 23 in ed81ce2

torch.manual_seed(args.seed)

[tanghengjian]

run with default value.
by the way, i found cifar10 dataset is only 32*32 pixel with 10 classes, it means the y condition changes from 0 to 9.
do you have tested the mscoco dataset in DiT model with label condition?

[zhengyu-su]

run with default value. by the way, i found cifar10 dataset is only 32*32 pixel with 10 classes, it means the y condition changes from 0 to 9. do you have tested the mscoco dataset in DiT model with label condition?

How do you link CIFAR10 classes to the ImageNet 1k classes?