May I ask how many iterations it takes for Stage 1 and Stage 2 respectively to converge on YouTube-VIS?

[12shuai]

[12shuai]

By the way, is it possible to train the model without Stage 1? Is the training able to converge?

[12shuai]

moreover, i find that the Instance-Enhancer is missing in the training&inference codes. is it necessary for SVD-version?

[pixeli99]

Hi, @12shuai
I want to confirm whether your basemodel is SVD or modelscope's T2V?

[12shuai]

SVD