@IcePrick he's known to misrepresent the condition of his films, and to take a very long time to ship them (probably to prevent buyers claiming a refund under ebay's buyer protection policy).
Well I'm not 100% certain, but I think that seller "salerare" works with "filmprint" who is definitely not to be trusted for reasons I won't go into in the shoutbox - I will just say I would highly suggest avoiding both sellers as well as the South African seller "video786".
Captcha is definitely useful, but people should abstain from Google's reCaptcha imo, it's pure evil. There are many captcha variants out there that should do for a small-medium size blog.
Latest year I received like thousand spam posts on my blog, and few dozens real posts... now, after including captcha, almost no spam at all... I hate captcha, but it is so useful!
I dont know all that much about machine learning, but from the limited experience I have with it, it needs to be trained on a representative data set. For example I don't think this would work very well as-is, because it's trained on losslessly downsampled images, whereas a lot of the content we might be interested in upsampling is A. compressed and B. already initially sampled at a resolution that is low. Photoshop has a ML-supported upscaling algorithm and it works the best on high quality (low compression) downsampled images. It doesn't work very well on images that are noticeably compressed and low resolution to begin with, it tends to just amplify the compression artifacts like blocking,.
Yeah, you can use ESRGAN, but it won't work very well with movies unless you get big data set with movie frames and you'll teach your neural network for long time.