How many steps do stage2 and stage3 trained?

[xilanhua12138]

No description provided.

[xilanhua12138]

@zsxkib @guozinan126 @ToTheBeginning

[ToTheBeginning]

For the XL model, each of stages 2 and 3 will not exceed 30,000 iterations. We mentioned in the paper that using only ID loss can significantly improve similarity, especially in terms of quantitative metrics, but it can lead to a reward-hacking issue, so the image quality will degrade more seriously after training for a long time.