I have a video I want to upscale, but frankly, I’m not sure what all is wrong with it, and what features would best help in producing a final end result.
And these are just old videos. Not super important. So, where I’d like them improved, I’d prefer to not have to test out a bunch of settings.
It’d be nice if you trained something that can judge whether a certain filter is going to modify the video in a way that most users are likely to like (human feedback), so that you can recommend the right settings automatically.
Sounds like you want Proteus Auto.
Proteus’ results did not turn out great when I tried it.
Plus, every model is going to have pros and cons.
In my tests one of the ones specialized for low quality had better results … It was old home videos that it would’ve been nice to restore.
Okay. Well, that’s the closest to doing what you want right now. As in: Artemis was the main set of models, but was very rigid on what movies it could give good results on. Proteus was made to be an adjustable replacement to the Artemis models. So in a lesser way, it is the outcome of your request.
I also have several old home videos that I want to enhance, and most of the models don’t seem to do anything. Iris does it’s trademark gimmicky face enhancement, but it doesn’t help the rest of the image.
This is a good idea, I just see it being added as a new model. I mention Proteus Auto, because that is the current state of their ability to guess the right output.
I think it’s hard to get one model to do many complex things at the same time. Plus, the model would end up bigger and slower.
So, kinda like in the opensource LLM space, they may be better off with some kind of chainable mixture of experts model, where each mini-model can make certain kinds of improvements.
And then you could try and improve a video by running it through multiple relevant AI based on what was found in the video (or maybe even just the scene).
… Of course, that would be a lot of work. But, if they want to win, then they need to be outright better.
