Because you want the A.i. (tvai_up) to be part of the process of upscaling the image! The way I had it was wrong:
tvai_up=model=prob-3:scale=0:w=1452:h=816
As that just does the A.i. on the only-cropped 1452x816p image (which then only gets upsaled with just a Lanczos resize in ffmpeg, later on). Not that I saw any difference in the output, tbh, but still.