Bad(interlaced, overlayed like) inference output of `rife-4.5` and `rife-4.6` models

[Justin62628]

Using the same script of inference implemented by nihui, the following models produce different outputs as followed: (uhd_mode = False and tta_mode = False)

• rife-v4_ensembleTrue_fastFalse or rife-v4_ensembleFalse_fastTrue , which is quite normal, while:

• rife-4.5 and rife-4.6, no matter ensemble is True or False , definitely wrong.

There might be a problem with the quantitization or conversion of the ensemble feature. Would it be convenient for you to share the scripts for model conversions(pytorch model to ncnn .params & .bins)? Nihui says she converted the rife-v4 model using pnnx, which could make a difference I guess.

[styler00dollar]

I don't use pnnx. I use self written code and workarounds to export. The model itself won't change anyway, only the param file. Please test with my vapoursynth code again instead of nihui. It worked fine there iirc. I use different ncnn submodule, but I don't think that is it.

[Justin62628]

uh you're right. lmao after some debug I found that using Intel GPU is the reason for that overlayed output. NVIDIA GPU works like a charm.