I have read the code of the "for loop" in "demo_gradio.py" from line 196 to 305 to generate video frame by frame.
I could not find any codes to distinguish from the first-frame generation to the others, which is important to the anti-shifting strategy described in the paper (the first frame generation should also generate the last two frames as condition for later generation).
I am confused where in the code implement this anti-shifting strategy and how?