Upscale low res card images to high res using Deep Learning

[mtgcard]

I can't send a link to the website yet.

[austinio7116]

Had a quick go at KHM https://drive.google.com/folderview?id= ... PwA7mH2HR8

[Arcanist]

Had a quick go at KHM https://drive.google.com/folderview?id= ... PwA7mH2HR8

Card names which contain the word "and" like "Battle of Frost and Fire" are systematically misnamed. Instead the card is named "Battle of Frostnd Fire" and missing two characters. That is something I noticed in previous archives as well but forgot to report. I assume that you are using a script to process the data and that those two characters get deleted for some reason?

[austinio7116]

Here is a first pass at STX - as some of you have noticed there is a systematic naming error with certain cases - not sure if it is me or the Wizard's website naming - but easy enough to correct:

https://drive.google.com/drive/folders/ ... sp=sharing

[scramsby]

These are amazing. I love the clear look of the upscales compared to the high res print scans. So much better for proxying/browsing.

I would love to get the full art lands from Unstable redone with your latest ESRGAN model, as the previous one I found is much worse than the latest releases you've put out. If it's less of a burden to cover older sets I'd be up for giving your model and code a spin to run it myself if you're up for sharing on GitHub or something.

Please let us know and thanks so much!

[austinio7116]

[scramsby]

Amazing, thank you!

[meep.o.matic]

Hey austino. Did you ever take a shot at old borddered (pre-8th) cards?

[meep.o.matic]

So... I've been training a model to upscale Gatherer pics as well and just wanted to stir back some interest in this project as well.
@austino7116 I'm mainly fumbling through this project, as I'm not a programer, but maybe we can openly discuss some technical stuff about AI upscaling. Right now I'm training another model that should look better than these, 115000 iterations in.

I was hoping to give foreign languages some love as well.

I encourage you to check the full resolution images as the previews I'm showing are just 1080p printscreens of the scaled image.
Damn - https://cdn.discordapp.com/attachments/547949806761410560/856604739671949312/Imassge.ashx-comparison.png
Delay - https://ibb.co/0Bxrh6x

[mtgcard]

Can you make a tutorial?
I want to learn!
@meep.o.matic

[AStrangerIsHere]

So... I've been training a model to upscale Gatherer pics as well and just wanted to stir back some interest in this project as well.

So what do you use to uspcale pictures? How do you do it? Do you need some particular knowledge?

[meep.o.matic]

I actually got started searching keywords and concepts from this thread. For the sake of paying it forward, I'll facilitate the info.

This is where you get started. Try to install the latest versions available for the torch, torchvision and torchaudio modules and not necessarily those that are in the tutorials:
https://upscale.wiki/wiki/ESRGAN_Installation_Guide_for_Windows
https://www.youtube.com/watch?v=w4Hb7tyDsWE&list=PLZpgUFpOvR7EG6RqFRxjhPUtILoiwuxgw

This is where you get started with training:
https://upscale.wiki/wiki/Beginner_Training_Guide
https://www.youtube.com/watch?v=KaMYsxWkmww

These are the repos I recomend using. They are being actively maintained. I've had success so far with these, although I haven't really tried others:
"Upscaler" - https://github.com/victorca25/iNNfer
BasicSR - https://github.com/victorca25/BasicSR

This is the current repository for public pre-trained models:
https://upscale.wiki/w/index.php?title=Model_Database&oldid=1194

This is the program I use to upscale more often because of the GUI. Never used Image Enhancer Enhancer Utility:
https://github.com/n00mkrad/cupscale

This is a useful program to help you prepare your datasets:
https://github.com/n00mkrad/magick-utils

Everything else you learn along the way.
It's not hard to follow along the tutorials and figure this stuff out.

[AStrangerIsHere]

I actually got started searching keywords and concepts from this thread. For the sake of paying it forward, I'll facilitate the info.

This is where you get started. Try to install the latest versions available for the torch, torchvision and torchaudio modules and not necessarily those that are in the tutorials:
https://upscale.wiki/wiki/ESRGAN_Installation_Guide_for_Windows
https://www.youtube.com/watch?v=w4Hb7tyDsWE&list=PLZpgUFpOvR7EG6RqFRxjhPUtILoiwuxgw

This is where you get started with training:
https://upscale.wiki/wiki/Beginner_Training_Guide
https://www.youtube.com/watch?v=KaMYsxWkmww

These are the repos I recomend using. They are being actively maintained. I've had success so far with these, although I haven't really tried others:
"Upscaler" - https://github.com/victorca25/iNNfer
BasicSR - https://github.com/victorca25/BasicSR

This is the current repository for public pre-trained models:
https://upscale.wiki/w/index.php?title=Model_Database&oldid=1194

This is the program I use to upscale more often because of the GUI. Never used Image Enhancer Enhancer Utility:
https://github.com/n00mkrad/cupscale

This is a useful program to help you prepare your datasets:
https://github.com/n00mkrad/magick-utils

Everything else you learn along the way.
It's not hard to follow along the tutorials and figure this stuff out.

Thank you for all the information, that's very useful! Also, is the training process long?

Edit: what model do you use?

[meep.o.matic]

Yes. The image I've posted is from the 150000 iteration checkpoint, which should be arround 1/4 of the expected training epochs to "completion". Took me 2 and 1/2 days with an RTX 3080 to get here.
The process speed is highly dependant on the settings and hardware. More available VRAM will let you push your settigns and allow for faster training.
Good dataset curation will also speed up good results, particularly for models with low variation between images.

[Barabba]

Hi guys, Modern Horizon 2 hq/Upscaler please? i wait responspe. Ty very much!