search

nihui / rife-ncnn-vulkan

RIFE, Real-Time Intermediate Flow Estimation for Video Frame Interpolation implemented with ncnn library
MIT License
822 stars 70 forks source link

[feature request] add support for RIFE v4.4 model #50

Closed hooke007 closed 2 years ago

hooke007 commented 2 years ago

Recently Practical-RIFE released their more v4 series models, could you spare some time to port them?

styler00dollar commented 2 years ago

I recently converted 4.3 and 4.4 (with and without ensembling/fastmode) for my custom vapoursynth extension here. They will probably work with her binary too.

hooke007 commented 2 years ago

glad to hear that, could you explain more about ensemble and fast's meaning in your fork?

styler00dollar commented 2 years ago

The rife architecture has two parameters called ensembling and fastmode integrated. You can't just export rife into one single model without such parameters. Nihui always just exports the fastest model (ensemble False, Fast True), while the other ones can result in a better image. Ensembling combines the flow of two different interpolations within the model and if fastmode is false it calls contextnet. You can see it here.