Hacker News Reader

Comment by fxtentacle | Hacker News Reader

This sounds like it could be a great inspiration for AI-based projects.

I understand that for their flagship production, they want people to draw true 4k all digital. But for most projects, I'd wager that a good AI upscaling and cleanup can take a hand-drawn paper character to a usable 4K plate.

Similarly, I don't think the contractors doing in-between frames need to work in HDR. It should be good enough to upconvert their images from SDR to HDR based on the adjacent HDR keyframes.

As for the background images, I myself already felt frustrated a few times when trying to get my 12MP SDR vacation photos displaying nicely on a 4K HDR TV, so solving that problem might even have a consumer angle.

View on HN

There is actually precedent for that. Waifu2x [1] is an image upscaler for anime-style art based on deep convolutional neural networks. As far as I know, it's mostly used by people who want to upscale fan art they found on the internet. It works pretty well, though, and it (or similar) could serve as a basis for the kind of workflow you're talking about.

[1] http://waifu2x.udp.jp/

There is also Anime4K [1]. It doesn't use machine learning though.

[1]: https://github.com/bloc97/Anime4K

waifu2x is not suited for actual anime almost at all - its upscaling is primarily trained on sharp lineart, and lineart in actual anime is generally quite soft.

Anime4K has even less going for it than waifu2x does, it's pretty much just simple warpsharping with a fancy marketing name.

Neither of these are particularly interesting as far as upscaling / remastering goes, and I'm saying that as someone who has done SD to HD anime upscaling/remastering projects from SD masters myself. And while you can get results that look pretty solid for the most part, it is ultimately still going to look like an upscale when compared to native footage. Here's a comparison with two examples:

https://slow.pics/c/wNZKk3WQ

As you can see in the second example, the upscaled nature of the SD to HD conversion is especially clear in areas where the lineart has been very tightly packed originally.

waifu2x is quite reasonable for manga though; it's quite commonly sharp lineart. I ran it on some old poor-quality jpg scanslations I had, and the results were good.

Whenever I see that name, I immediately think “oh the poorly rebranded version of NNEDI3”.

I'd say it's well branded, for the target demographic of weebs. It's also a generally memorable name.

Weebs are not the only target demographic of scalers, though. They simply lifted the work of others to associate decent image scaling with anime fandom. Importantly though, scalers shouldn't need a brand since they are not a product. They are a tool that can be implemented, and to that end "neural net edge detection" is much more informative and less likely to turn off an audience than "waifu2x". waifu2x is a product name. All it tells me is something about the publisher and their target audience.

> All it tells me is something about the publisher and their target audience.

Yeah, that their demographic is weebs... I'm not sure what your point is, but mine is that waifu2x is effectively branded, not poorly branded.

“Effectively branded” would be branding for the audience that are ripe for using the tool: almost everyone who watches video and has a high resolution screen.

I'm pretty confident that nagadomi intends for waifu2x to be used for anime fanart (they say as much) and for that, it is effectively branded.

Just because something wasn't directed at a general audience doesn't mean it was poorly branded.

waifu2x is optimized for anime.

Because of its anime upscaling success, it gained a lot of awareness and is now used in more non-photorealistic images in general, with a very decent popularity.

It was used that way before it was named waifu2x. It’s just NNEDI3.

Reply on news.ycombinator.com