Did anyone notice this (new advancements in texture upscaling)

[SCO]

 

It seems like the new methods of texture upscaling with neural networks are significantly better at extrapolating details than the previous geometric methods. I'd love if the enhanced editions got this applied to their area and npc textures (with some QA ofc). Especially because of the zoom in.

 

These comparisons work better if you zoom in to see what happens at non-standard zooms (like the EE).

(original)

vs

Edited January 13, 2019 by SCO

[lynx]

Very nice results indeed! Hopefully someone will integrate it into SDL, so we geet more upscaling options for free.

[SCO]

It's not a runtime method. That's why i mentioned EE. Or gemrb, now that i think about it, but i don't think gemrb has a zoom in function yet.

This method appears to work better if the image is already reasonably detailed and 'true life' related (because the extrapolations attempt to match textures). If it's small enough resolution or 'artistic' enough or both the results can be disastrous:

But if it's 'orderly' enough, they appear to work great (i wish scummvm and dosbox had generic texture replacement):

http://screenshotcomparison.com/comparison/127742

(this is a longest journey background).

 

Plenty of mid-late 90's and early 2000 games that had blurry textures even on their native resolution that could be made much sharper with this effort without source code access if these two projects had texture replacement capability.

Though maybe it could be runtime if you're willing to waste a lot of memory and a little bit of time if you have the hardware and the training set right there. It apparently takes 10 seconds on a 6gb GPU or something like that. I agree that would be a coup, especially since people wouldn't have to download mods or wait for companies to decide on re-releases.

Edited January 13, 2019 by SCO

[lynx]

Too bad, this means it would also complicate distribution if you wanted to do it in advance. It would basically have to be a mod for us, each user doing it themselves.

[SCO]

Yeah. I also notice that even on 'good' results, the NN can get confused about what material some of the textures are supposed to be. I think this sort of stuff will have to lean heavily in the training sets and many unusual environments or perspectives or multiple depth scenes might get screwed by a one size fits all approach.

 

For instance on that longest journey screenshot above, the bricks and vine on that building on the background appear not to be very brick or vine like on one side of the wall, then the wall turns into obvious bricks again on the other. I'm actually surprised the Baldur's gate screenshots turned so well, but the guy says he used a different method::

Quote

I'm using the RRDB_PSNR_x4 model as the RRDB_ESRGAN_x4 introduces a lot of artifact.

 

Edited January 13, 2019 by SCO

[SCO]

BTW, here is the paper and repository:

https://github.com/xinntao/ESRGAN