QualityScaler: Image/video deeplearning upscaler with any GPU | Hacker News Reader

QualityScaler: Image/video deeplearning upscaler with any GPU

github.com

99 pointsjonifico21 comments

For anyone interested in different upscaling models, there is a collection in the wiki here:

https://upscale.wiki/wiki/Model_Database

My personal favourite is Remacri for photographs, but you should download a few of them and try what works best for the photo/image you have.

You can use either Automatic1111's repo for stable diffusion to try them out (put the model file in the ESRGAN directory), or also chaiNNer (https://github.com/chaiNNer-org/chaiNNer)

A VERY similar project was submitted to Awesome F/OSS, Upscayl:

https://github.com/upscayl/upscayl

Wonder what the comparison is…

wondering the same

Upscayl seems to have somer specific hw requirements (says it requires a GPU and won't work on most iGPUs).

Thanks for digging this out — looks like it was kinda in the tagline —- “any GPU”

Upscayl works fine, albeit exceptionally slowly, on my Ryzen APU (5600G). It uses the GPU, it's just that the GPU is very small.

I suspect they are just not really wanting to advertise such a slow user experience, which is reasonable.

Windows Defender says that the binary has Trojan:Win32/Wacatac.H!ml inside it.

Apparently it's a known issue: https://github.com/Djdefrag/QualityScaler/issues/21

Well https://www.virustotal.com/gui/file/fd86f7f9fa0b8f584e0becc4...

This looks fancy. Any chance of anyone implementing this as a ffmpeg filter?

How does this compare to Real-ESRGAN?

Is it better than waifu2x?

Nothing I've tried is. I might try this one out to see, but if anyone who has can comment on how it compares, I'd appreciate it.

Most modern architectures are better than waifu2x, this includes real-ESRGAN trained on the anime dataset, SwinIR.

I'll give it a go, but I've tried several and waifu2x gave significantly better results, so I'm not expecting much will be happy to be wrong!

My use case is 2x upscaling detailed, very small pixel-art-style images, like icons and retro game assets.

What kind of input do you need?

I'm getting poor results trying with some stamps I've scanned and some photos that were lower resolution than I liked.

Or does it need to be studio quality footage?

It looks pretty good, but especially with the Spiderman comparison you can notice a little smudging and it basically loses the pattern on the suit.

Looks to be limited to running on Windows only.

Anyone found something like this for audio?

You’d need to be more specific. If you’re talking about taking CD quality audio and Up-sampling to “high res” audio, what you’re actually doing is fabricating new inaudible frequencies which do nothing except potentially degrade the audible sound if your tweeters aren’t up to the task of reproducing these useless frequencies perfectly.

What I am excited for is a narrowband to wideband algorithm which can turn a Ye Olde Microphone recording or possibly even a telephone conversation into something resembling a modern recording. Obviously there’s only so much you can do with a recording made on a terrible microphone, but it should be possible to make historical recordings clearer.

Are any open source image denoisers available ?

There's waifu2x, it does scaling but also denoising:

https://github.com/nagadomi/waifu2x

And other implementations thereof, e.g:

https://github.com/nihui/waifu2x-ncnn-vulkan

Reply on news.ycombinator.com