My scenario is this: I have a DVD, and a blu-ray (nice quality, 1080p)
The DVD is basically a music performance in a club. The blu-ray is of a film that uses many, but by no means all of the exact scenes from that performance. I want to upscale the DVD - and the results I’m getting with VEAI are … decent, but I wish I could improve on them (and I’ve tried a lot of different methods and model combinations etc)
Is there any way of training / tailoring or customising an existing model with the SD / HD versions on the DVD / blu-ray, to get a model suitable for upscaling the whole DVD? Or would I need to go outside of VEAI for that (I’m comfortable with coding and exploring new territories, but right now, I don’t really have a starting point)
Take a look at ESRGAN if you want to train your own models. It’s slow but I believe there’s quite a few variations of it that attempt to increase performance. ESRGAN Plus and some old attempts at converting it to utilize Tensor Flow are what come to mind. But I don’t have the knowledge to actually look into the forks.
It’s popular in upscaling or cleaning up game textures while also being used on old animations since Topaz doesn’t have a good enough solution for those.
But that doesn’t mean people haven’t created models for lifelike scenes. It’s just not as common.
Training ESRGAN is probably only through command. But in utilizing it I could recommend Cupscale as a GUI option.
All of which are available via Github.
Thanks. I’ve been looking at ESRGAN, and coming to similar conclusions - although as a Mac user, I’ll likely have to go the Colab route. Just trying to find a good starting point.
