Questions about the second(pretrain) and third(finetune) stage training?

[dydxdt]

Great job! I have some questions about training. And I want to train stage 2 and 3 using my own dataset.

As the above screenshot shows, you just give one command for training. I also read PixArt_xl2_internal.py and PixArt_xl2_sam.py. I'm not very sure that the former is used for the third stage training and the latter for second stage, is that right? Can you specify more about the difference of parameter setting in stage 2 and 3?
And for PixArt_xl2_img512_internal.py and PixArt_xl2_img1024_internal.py, they both can be used in the third stage training?(Maybe as mentioned in the paper the internal dataset you used in the third stage are not open sourced). Does it mean that choosing 512 or 1024 depends on my situation?
Really hope for your reply! Thanks

[dydxdt]

Besides, in the third stage, it's better to use "InternalDataMS" for multi-scale training, is that right?