Incorporate newer models, such as 2x-AnimeSharpV4

[thekryz]

I've been looking around for better upscaling models and come across 2x-AnimeSharpV4 by Kim3091. I couldn't find a direct comparison with e.g. animevideo-v3 / Real-ESRGAN, but it looks very good. I'd love to see more models such as this incorporated into video2x.

Alternatively, providing visual comparisons in case these models are not worth incorporating would be great.

[k4yt3x]

Here's the issue with model requests like this: #1327 (comment)

I.e., how do we devise a standard to evaluate which kinds of models should be added and which ones not? Merely based on if someone suggested it seems insufficient.

Since Real-ESRGAN should be able to run these models once you've converted them into ncnn format, it might be a better idea to add support for automatically searching for models in the models directory so users can add any custom models?

[thekryz]

I.e., how do we devise a standard to evaluate which kinds of models should be added and which ones not? Merely based on if someone suggested it seems insufficient.

Since Real-ESRGAN should be able to run these models once you've converted them into ncnn format, it might be a better idea to add support for automatically searching for models in the models directory so users can add any custom models?

I absolutely see your point. And I appreciate your suggestion for custom models and would love that to be possible!

Thinking a bit further: Perhaps it would be possible to create a benchmark for upscaling with a few pictures and short video clips, that can be thrown at different models and added to a results overview? The base images for this benchmark could maybe even be created with AI, so they're not copyright-protected. If that's an option, I'd be happy to have AI create some if needed.

[k4yt3x]

We can probably downscale a test video and use something like VMAF to compare the quality of the upscaled video to the ground truth at the same resolution.

It might also be easier to support loading arbitrary NCNN models so users can try to convert the models and toss in whatever they want. I'll just need to write the logic to search for custom model files.