Accelerated reasoning process

[lgthappy]

Hello,I am using the 4.18lite version of the model and have set fp16 and scale=0.5 but it's not as fast as I would like, is there any other way to speed it up?

[hzwer]

Hello, there is no solution for now. I am considering rolling out a faster version, but at the cost of reduced effectiveness.

[TNTwise]

You can use rife v2, that helps somewhat https://github.com/AmusementClub/vs-mlrt/releases/tag/external-models, if you don't want to use tensorrt, you can use vs rife arch's https://github.com/HolyWu/vs-rife, which increase performance somewhat by caching back-warp and tenflow. I hacked rifev2 into an optimized CUDA version, but it really only works with https://github.com/TNTwise/REAL-Video-Enhancer alpha

If it's not increasing performance, you are encode limited, which is hard to fix unless you use nvenc but it's not that good IMO.