Does CogVideoX1.5-5B I2V support Portrait instead of Landscape?
Support, his resolution is free, you can set it to portrait as long as the width is not less than 768.
I found that if I change height larger than width, then I have to swap sample_height and sample_width in transformer/config.json or errors will occur in the rope positional embedding part.
I think the reason is that the code logic of rope positional embedding assumes that sample_width is larger than sample_height, which is set to 170 and 96 respectively. After I swap sample_height and sample_width, there will be no errors for portrait inference.
I am not sure this will harm the generation results or not.