Questions about reproducing Stage I&II and published weights

[k-zha14]

Hi, thanks for your brilliant works and kindly open-source efforts. I'm puzzled by two questions and hoping for your replay.

1. Could you explain the differences between public pre-trained weights (in red box)? Do they have some relationships with Tab. 4 in the Pixart-alpha paper?
   

2. In Sec 2.2, you illustrated the whole training is deivided into THREE stages. As for Stage I-Pixel dependency learning, I think you initial the model with DiT weights to boost efficiency, called reparameterization. In other words, there is no training need on stage I. However, I find that there are still 300K steps on ImageNet in Table 4 (red box). So, do I miss some details? And what is the training obejective of the 300K steps ImageNet training? text2img or cls2img?
   

Thanks you so mush.