The version of CogVideoX.

[LeoniusChen]

Great work! I would like to know which version of CogVideoX is adopted for the backbone of this work. CogVideoX1.0-t2v, CogVideoX1.0-i2v, CogVideoX1.5-t2v or CogVideoX1.5-i2v?

[SHYuanBest]

Thanks for you interest, the current open-source version of ConsisID is based on CogVideoX1.0-i2v. We have another version based on CogVideoX1.5-t2v and CogVideoX1.5-i2v that haven't be open-sourced.

[LeoniusChen]

I have one more question: why was the i2v model backbone chosen for the experiments in the released code? This task seems to be more related to t2v. In your non-open-sourced models, what characteristics would the t2v backbone have? For example, would it support higher dynamics but worse identity preservation?

[SHYuanBest]

i2v will converge faster, but the performance will be worse than t2v