How to extract t5 or video ae feature to fit T2V_Feature_Dataset？

[knightyxp]

Hi, I noticed that pre-extracting features for 512-frame videos can significantly reduce memory usage. Could you please share the code for pre-computing video and T5 features? Additionally, the memory savings from compress-kv are limited. Could this be due to the compress_kv factor being set to 1? Would higher compression factors like 2 or 4 further reduce memory usage? What are the potential fault?