Request for the training configs of DiT-XL

[a-little-hoof]

Hi there, I really like your work!

I noticed that there's only the config of DiTDH under configs/stage2/training/ImageNet256/DiTDH-S_DINOv2-B.yaml. I wonder whether you have plans to share the training config of LighteningDiT?

We are using the updated codebase with settings mentioned in the paper and the pretrained decoder you have provided, but fail to reproduce gfid（e.g for SigLIPv2 we have around 5.7， which is higher than what has been reported by the paper)

This is the setting from the paper:

And this is the exact yaml file we are using for training lighteningDIT:

DiTXL_SigLip2-B.yaml

[bytetriper]

Hi,

The training config seems pretty aligned. Can you share the inference setup you use? Maybe try different bs and fp32 for inference.

[a-little-hoof]

Hi,

Thanks for the quick reply. Happy new year!

This is the inference setup we are using:

DiTXL_SigLip2-B.yaml

Could you please take a quick look and let me know if anything stands out as inconsistent with the intended inference settings?