To be painfully honest, I’m not a fan of upscaling or motion interpolation. That video looks good, and there are legitimate uses for both those things, but the whole AI enhancemet craze is really driving me nuts (it’s not your fault @junker).
AI still has a ton of problems. For one, using an image AI upscaler for video often can produce results that are not consistent between frames. Second, AI has a problem with generalization; AI is only as good as the data it’s trained on. For example, if I train an AI to upscale realistic images but I give it cartoon images, the AI will try putting realistic detail on cartoon images which will produce unnatural and dissatisfying results. There’s also the heavy performance cost and heavy memory usage (and even way more if you want to train one yourself). And there’s the problem of (at least for restoration purists) “was this extracted detail actually recorded in the image or hallucinated by the AI?” Explainability is, in fact, one of the biggest problems with all of AI in general.
I would imagine that most of these problems will get diminished as further advances are made (the activity going on in that field is ridiculously high). But for now, these problems are still pretty big problems.
At least they used DAIN for motion interpolation, which I’ve read about months ago and is definitely one of the best algorithms of its kind out there.
And at least they didn’t use DeepRemaster; the algorithm has a ton of practical problems because it tries to be an algorithm that does everything (which is a definitive path to failure and suffering).
I’ve been conceptualizing an ideal AI algorithm that wouldn’t have some of the problems of typical AI algorithms, but I always have a full plate with projects and stuff. Plus, I don’t have good enough hardware.