update will suitable for uoload pretrained model

[renrenzsbbb]

Thanks for your great work. I find that thers are some commit to update the i2vdataset, like add image transform to it and remove the reference image as the first image . I argue that these two change should be consistent with your trainning mode so that there will be a congruent new pretrained model. but unfortunately I can't find it. So I wonder that the pretrained model is suitable for new change like above mentioned.

[zRzRzRzRzRzRzR]

These are because we lacked alignment with the training team in our previous implementation.
Firstly, the length of the training team's training videos is 8N + 1 or 16N +. That is, 81 frames / 49 frames of images, not 48 frames of images.
Therefore, when preparing the dataset, each complete dataset should be 8N + 1, not by appending the first image to 8N.