Page 6 of 19

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 21 Sep 2017, 21:38
by Huggybaby
This project is amazing! These images are light years ahead of what we had in the old days.

austinio7116, please update your first post with a current before and after shot. Also leave up the pics you have now so we can see the progress you've achieved.

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 22 Sep 2017, 10:19
by MaxMtg
I tried to use Waifu 2x to better the image quality of CMA low-res pictures but they came vastly inferior to your image quality

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 22 Sep 2017, 13:06
by odino182
austinio7116 can you this amazing work for "stardand" italian card? ty alot

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 20 Oct 2017, 20:02
by austinio7116
MaxMtg wrote:I tried to use Waifu 2x to better the image quality of CMA low-res pictures but they came vastly inferior to your image quality
Yes, I think my final results are better for magic cards than Waifu, although I do think Waifu is great for anime art.

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 22 Oct 2017, 06:16
by austinio7116
Here's the low res and upsampled versions of one of the nicer results on the art from XLN:

Emperor's Vanguard.full (2).jpg

Emperor's Vanguard.full.png.xhlq (2).jpg

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 27 Oct 2017, 22:21
by austinio7116
Here are some sets of images showing the low res, the super resolution - and now the newly released scans for comparison:

compare1.jpg

compare2.jpg

compare3.jpg

compare4.jpg

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 28 Oct 2017, 02:53
by kevlahnota
Nice comparison. IMHO, if you are using mobile device not greater than 9" display, the super resolution will do the job (also the file size is very friendly for mobile devices). The card scans is much suited to PC with big monitors...

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 28 Oct 2017, 12:34
by austinio7116
Here is a comparison with other upsampling methods:

The pictures are: Low Res, Bicubic interpolation, Waifu 2x (just the standard web version run twice - not specially trained on Magic cards), My DCGAN, HQ scan

compare5-1.jpg

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 31 Oct 2017, 02:20
by Ayssia
Any Iconic Masters update?

Edit: nvm... found it, sry.

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 02 Nov 2017, 11:14
by MaxMtg
Can you upscale Duel Decks goblins vs merfolk?

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 02 Nov 2017, 11:27
by austinio7116
Yes, I'll have a look at those.

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 02 Nov 2017, 22:58
by austinio7116

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 03 Nov 2017, 00:06
by MaxMtg
austinio7116 wrote:https://drive.google.com/drive/folders/1gnO6dR8Tjlu3IhoxIMu5kvCeMzEYaGbL?usp=sharing

lowres


highres
It's amazing how those sometimes look even better than the scanned cards
Have you done this to commander anthology?

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 03 Nov 2017, 06:31
by austinio7116
I think the colour depth, vibrance and colour consistency can be better with this method as printing/scanning is bound to reduce this a little, however the detail and sharpness can never be as good as the scans at 300dpi.

If I can find 370x265 CMA images from Wizards I should be able to upsample them.

Re: Upscale low res card images to high res using Deep Learn

PostPosted: 03 Nov 2017, 09:08
by austinio7116
I can't see 370x265 images, only lower Res gatherer ones, perhaps that's why your Waifu upsampling was so noticeably inferior. I can try upsampling the gatherer images but would not expect to get such good results, the network is specifically trained on, and so optimised for 370x265 input.