Inference Time

[Platorius]

I love this list and i like it, that you go for LPIPS

But i would suggest one metric more: Inference time which it took for one frame to generate (with a model the researchers trained).

Because doesn't matter how good sth is, if you get 0.5FPS output-speed on a RTX 4090 it is basically useless. Even in 10 years this would be still to slow. So important is quality Improvement while still beeing fast. Good quality but super slow will result in "no one uses that". It has no practical meaning.

So that's my suggestion: Trying to use this values in your tables. I know sadly most researchers do not tell, how fast it is.

And i know it would need also the used resolution and graphcis card, but it could be said too.

But it is only a suggestion.

Good list. And i am also for using LPIPS only and going away from PSNR and SSIM.