Upscaling images of Pokemon cards using AI

I haven’t seen any discussions of AI being used to upscale images of Pokemon cards. I’ve tried a few different image sources and types, and thought it was worth sharing.
The software i’ve used is Topaz Labs Gigapixel AI.
For those whom are unaware of what it means to upscale something, it essentially means to improve the size or quality of a said media. In this instance I am using various source images, from low res card scans (presumably) to pictures of PSA slabs. The application will be increasing both the image size and quality of source images.
As we go through the images I might comment on the performance of the AI. My use of this software is purely to satisfy my curiosity.

1st edition Charizard
Source: pkmncards

Original | Upscaled (4x)

I think this is pretty impressive. Some of the artifacting is carrying through as design elements which isn’t ideal, but it still shows a solid result.


Lucky Stadium Trainer CardSource: pkmncards

Original | Upscaled (4x)

Wow. The characters really pop out in the upscaled version as the blur has been removed. The buildings seem to have lost some of their windows however on the buildings lol.

The AI isn’t quite sure what to make of the windows.

Misty’s Cerulean City Gym Trainer CardSoure: pkmncards

Original | Upscale (4x)

The upscaled repaint by the AI has added clarity and to me this appears to contribute to more depth.

The artwork in the original image is certainly softer. The water is perfectly recreated by the AI at a higher resolution.

Spiky-eared PichuSource: Pinterest

Original | Upscale (x4)

I was very curious to see how the AI would handle the colouring lines and to be honest I think it did an amazing job with its interpretation.The AI has resorted to a heavy use of spaced out dots however to achieve the gradient effect in the titlebox and border, and this could be seen as an over refinement.


But that clarity and removal of blur from the source image is staggering.

**Pikachu Illustrator (PSA slab)**Source: ebay (smpratte)

Original | Upscaled (2x is hosted (due to imgur filesize restrictions))

Again, we see an over refinement of the gradient in some areas.


The Pikachu component of the card is cleaner looking although it still shows a harsh treatment of the gradients.


Misty & Pikachu (PSA Slab)
Source: ebay (PWCC)
Original | Upscale (4x)
The line work has been repainted beautifully. The fill in the mouth and the shadow under the arm is a little odd, but overall the AI achieved an impressive result.


I can’t unsee pikachu having claws now.

But let’s try something different now and look at an image scan other than of a card.
A scan of Brock
Source: hirespokemon:
Original | Upscale (4x)

I’m quite impressed. Not that this is a particularly complex drawing, but still it performed better than I expected.

But enough of my comments. What do you all think about the technology?

12 Likes

This is super cool! I’m in to see more

1 Like

Wow this is really cool! The first application I could think of for something like this would be for the card mentioned in this thread made by @pichufan

Which Pokémon appear in Japan’s 019/P Pokémon Center Tokyo?

As you can see there’s a lot of detail in the card so I’m wondering how well the AI could upscale it to potentially tell us all of what’s in it.

Thanks for sharing this! I’m sure other members will think of cool applications for it.

1 Like

Instagram Filter for Cards more or less. Would be interested to see what it would do with a super high-res image with crisp ink dots. Try one of CGC’s pictures? They are some of the highest res photos I have seen, their paid ones more so.

Very cool! If anyone’s looking for a more open source solution there is also Waifu 2x which has been around for a bit waifu2x.udp.jp/

1 Like

Source

This is the best I could do. (2x upscale)

The thing is I had to reduce the rosette pattern to a large extent for the AI to know what I wanted it to focus on, otherwise it tries to clarify and focus on them.
Here is what it looks like with the rosette pattern in tact.
I did the removal in photoshop.
Also: I think one of the plush dolls is either an Omanyte or Omastar. The other might be a very poorly proportioned Ampharos.

1 Like

Yes, I would also be curious how it does with a better scan. The examples are pretty compressed which causes many of the artifacts and I wonder what it could really achieve by cleaning up and sharpen an already good image.

The problem is really the rosette patterns when you get close enough. If they are reduced then it’s ok to an extent otherwise it just wants to perfect and enhance what it sees.
I think I posted just before your response, but you can see in my previous post a decent example.

I’m thinking more of images like these:www.pojo.com/wp-content/uploads/2020/06/2003-Pokemon-Skyridge-Holo-Charizard-146-PSA-10-GEM-MINT.jpgassets.pokemon.com/assets/cms2/img/cards/web/SM5/SM5_EN_107.png
Not quite a 600dpi scan with the rosette but rather a mostly clean image with minimal compression artifacts that just doesn’t have sharp details when zoomed in. I think that’s why the Brock example you posted works well

Oh ok. Well here is the result.

Edit: it’s 4x and it shows the deficiencies with the text distortions most notably.

It does a good job at sharpening the image for sure. I was thinking about applications for archival purposes/generating a database of high quality/sharp artwork but it seems to alter the original work just a bit too much something like that. It would make a killer thumbnail for a youtube video though

At the moment it really depends on:a) the artwork and its complexities;b) the quality of the source images; and
c) the final intended use.
Currently it isn’t consistent enough for archival purposes because of the reliance on these variables. But when using adequate source images in combination with art containing elements it can identify, it does a fantastic job a lot of the time. I’m personally excited to see how rapidly the AI improves, as it inevitably will.

This is fantastic! There’s something deeply satisfying about seeing these images this crisp.

1 Like

Enhance! Enhance! Enhance!

This would be a cool way to scale up postcards to convert them into posters or playmats.

1 Like

This is like that stuff they always do=id in crime series from the 90s, enhance a 4 pixel photo untill you recognize a person on the picture in full HD. Works very well on the cards, I’m impressed!

Here is an example of 2 that are just shy of a 4x dimension increase. I could of course decrease the size after upscaling by compressing them but it would reflect on the final result unfairly compared to a direct output. But I will add one such example at the end.

The following postcards were sourced from seiry-pkmn.

01
original | upscaled02
original | upscaled

Here is a compressed example from hirespokemonoriginal | upscaled (4x, compressed to 20mb lossy JPEG from 87mb PNG)

Sorry for largely leaving out the side by side comparisons. I’m currently vacillating between a handful of projects at the moment.

3 Likes