How is the TCD method implemented?

[LJungang]

Hi, how is the TCD method implemented? I could not find it in the code.

Could you provide more detail about TCD or release the code,
thanks!

[Stevetich]

Hi, thanks for your attention to our work! We only provide the benchmark in this repository due to the overly tight schedule. However, I can provide more details for you before we upload the implementation of our TCD. Basically, we follow VCD to implement our TCD (I notice that they have fixed the bug in their method). The main difference is that we used model attributes to introduce the CD input, instead of using prepare_input_cd in VCD. This operation helps us to ignore the specific structure of the model, which can be used in different VideoLLMs.