>>8268868
There's so many factors involving this so I can't give a solid estimate without making a LOT of assumptions.
With my setup using a Ryzen 9 5900x and an RTX 3090, assuming 720p or 1080p 23.976fps input at 16:9 aspect ratio, using TensorRT through VideoJaNai, being able to encode while you upscale, using 2 models at the same time, it takes generally 0.404x of realtime, so basically if the video is 20 minutes, it would take 50 minutes to upscale.
If I'm running too low on RAM or doing other shit in the background, it can drop down to 0.379x or lower, just depends what you're doing.
If your video input is lower/higher resolution or you downscale to something higher than 720p, it can take longer, aswell as if you upscale to higher than I do, at 1440p.
Depending on your CPU and your encode parameters, the CPU could bottleneck you aswell.
Using different models also makes an impact on speed, and if you want speed, you could just run the ESRGAN model alone and it would probably be close to realtime, if not faster.
It really doesn't take long to upscale videos in 2024 if you have a decent setup.