Would love to get some tips!!!

If you used the model i would like to hear your feedbacks onit, it is only trained on faceforensics++ and celebDF-V2 datasets so the types of information on deepfakes is limited considering that now AI generated videos exist. I would love to get some tips and resources that i can probably use to make it better. I am aware the streamlit implementation is not upto the mark i wanted it so the user could select a specific moment on the uploaded video to detect deepfake but couldn't complete it int time. I am currently trying to learn and build some other projects but would be building a another version of this with improved architecture and datasets pretty soon.