AI enhanced textures for old games

[Dedicated_Dark]

Can't make it work. problem with torch module for python. For some reason python doesn't seem to find it and throws up that i don't have it where i just installed it via pip

Got the exact same shit. Great!

Edit: Bah! CPU is not working out. Left it for 3 hours and still no result. Shouldn't have tried with my pos AMD Laptop I guess. -_-

Edit: Runs when using network interpolation. Results are Amazing.

Edit: Great! Now my system hangs on file creation, fckn great. Ran twice properly. Fck it. I am done.

 

[Shinji]

More FF9:

These are not enhanced images -- look at the watermark on the bottom left corner -- these are the original renders made by the artist that worked on the game.

Taken straight from the guy's portfolio:
https://www.jakerowell.com/blog/

 

[sullynathan]

some of these look like paintings

 

[Zed Duke of Banville]

These are not enhanced images -- look at the watermark on the bottom left corner -- these are the original renders made by the artist that worked on the game.

Taken straight from the guy's portfolio:
https://www.jakerowell.com/blog/

And quite well-done, too.

Spoiler

 

[Perkel]

That would explain why they are blurry compared to others.
Still that FF9 shot in op is from AI stuff.

Don't think i will have time to figure out how to make it yourself today.
Whole christmass thing doesn't give me time to scratch my balls even.

 

[Dedicated_Dark]

So got it working. Installed on work system. Now the problem is I can't seem to run the same image through multiple passes, it seems to not produce a result when I put the image again for network interpolation. Any ideas?

 

[Perkel]

trying to get it working atm.

someone work on gothic:

 

[Perkel]

Got it working !
I tried earlier to use Cuda 9.0 and Python 3.6.6 and despite following and installing dependacies it still had problems with importing one module.
This time i installed Cuda10.0 and latest python 3 and no problems got it working at first time.

I think the comic jpg in test is pretty fucking amazing way to show how fucking good this algorithm is:
I upscaled initial pick from 62x90 to 500 so everyone could see better:

That is fucking crazy. Obviously there are some weird details with fingers but overall you would barely notice it at first glance.
Also there seems to be even better stuff coming in very short time. DNSR version will have build in method to properly handle downscalled images like case of textures.
Meaning that it should produce way less artifacts because it will know from start that images has been down-scaled in first place.

 

[Berekän]

I'm calling the cops.

 

[Rahdulan]

trying to get it working atm.

someone work on gothic:

/Gothic cut

Hard to complain to be honest. It just looks plain sharper without some modder's personal idea of what texture replacements should look like and alter overall aesthetic of the game. Nice.

That old ENHANCE nonsense from CSI shows is apparently becoming real in front of our eyes.

 

[Grotesque]

this this type of texture enhancement brings a certain flatness to it, it gives a pastel painting vibe

ok, this is fuckin freaky
you can't squeeze out that much information from the pixel bunch from the left without giving the AI additional information.
is the end result always the same? because fewer the initial information, the more variety the end result should have.

 

[†††]

Call me old fashioned but I like my old games to be pixelated.

You're not old-fashioned; if you were you would know that old games in their original resolution in CRT screens don't look pixelated. Playing Final Fantasy VII at 1080p on a flat screen is not how it was supposed to look like.

 

[Perkel]

this is how they are supposed to look on actual crt, and those are closeups,
scanlines effectively makes our brains add more gradient between pixels.

this this type of texture enhancement bring a certain flatness to it, it gives a pastel painting vibe

ok, this is fuckin freaky
you can't squeeze out that much information from the pixel bunch from the left without giving the AI additional information.
is the end result always the same? because fewer the initial information, the more variety the end result should have.

Yeah like how the fuck do you take so much detail out of like 50 pixel wide picture.
Yeah it creates additional details but we are talking here about really freaking small pictures so i doubt original artist would complain if he did see that. This depends on a training but apparently it learns how pictures are usually made and what should go after what.

BTW what settings did you use ? Trying atm planescape stuff but i can't make it work properly as it gives me more or less slight sharpening. I think

 

[Grotesque]

BTW what settings did you use ?

Diablo related image was taken from the internet

 

[Perkel]

Then probably it is not ERGAN but nvidia one from closed beta.
There are few algorythms out there.

ERGAN is for photo/paintings which were mostly downscaled via bicubic
SFRGAN which is similar but works better on art
Nvidia which seems to target everything but closed beta atm
Waifu2X which is specifically working for manga/anime

Nvidia one seems to be the most advanced as this will be made available to devs and mainly as a tool to upgrade textures.
Not only it will upscale but it will be able to morph realistically textures so for example wall on side of the road will have unique textures for whole streatch autogenerated from few textures.

 

[Wirdschowerdn]

Call me old fashioned but I like my old games to be pixelated.

You're not old-fashioned; if you were you would know that old games in their original resolution in CRT screens don't look pixelated. Playing Final Fantasy VII at 1080p on a flat screen is not how it was supposed to look like.

I know that. Perhaps I should have said "Low resolution" rather than pixelated. Sloppy wording.

 

[Dedicated_Dark]

You are going to have to open all these images in new tabs to see clearly. And then maybe zoom in as well.

^Original image, original resolution was around 800x600, upscaled to show the difference.

^1st pass. ESRGAN x4. Excellent results.

On subsequent passes ESRGAN results look the image was ripped apart at the same time noise was added.

^ESRGAN 3rd pass: Notice the bricks near the right wing, a fckn mess.

^2 ESRGAN and 1 Manga(Taken from the blog). Seems to have made the image look like a painting.

So far, maybe running through the sharpening(ESRGAN), then painting(Manga), then maybe a minor network interpolation at 0.6-8 should give decent results, I think. Possibly will have to edit the models itself to get decent results.

As for Gothic, it just seems to indefinitively sharpen the image, it looks rubbish. Don't think there is any point in doing this unless you directly run it through the textures files.

It's not as revolutionary as I thought but the first pass did produce a significant enough improvement over the original upscaled. The example images seem to produce truly incredible results.. .. might be something about the format of the images itself or other problems, doesn't seem to work well for me.

 

[DeepOcean]

I prefer pixelated, wake me up when they invent an Ai to make early 3d look good.

 

[Atrachasis]

Was the original image part of the training set?

 

[Valky]

Soon, the superior race will take the mantle of civilization.

 

[Dedicated_Dark]

Got it working !
I tried earlier to use Cuda 9.0 and Python 3.6.6 and despite following and installing dependacies it still had problems with importing one module.
This time i installed Cuda10.0 and latest python 3 and no problems got it working at first time.

I think the comic jpg in test is pretty fucking amazing way to show how fucking good this algorithm is:
I upscaled initial pick from 62x90 to 500 so everyone could see better:

That is fucking crazy. Obviously there are some weird details with fingers but overall you would barely notice it at first glance.
Also there seems to be even better stuff coming in very short time. DNSR version will have build in method to properly handle downscalled images like case of textures.
Meaning that it should produce way less artifacts because it will know from start that images has been down-scaled in first place.

So now, save that 1st image run it through ESRGAN again and show me the result.
Then try downscaling the 1st image back to it's original size, run it through ESRGAN again and show me the result.

Cause when I did the above, what I got is either a rubbish looking mess or pixelated garbage that looked the same the as the original image. I need to know if the images are a scam!

 

[Dedicated_Dark]

Original

ESRGAN

Seems to work well for art.

As a bonus here is Waifu4x result.

I am done. Will research into this shit when I actually need to use it, kind of annoying to randomly keep testing.

 

[FeelTheRads]

Excellent results

You're not old-fashioned; if you were you would know that old games in their original resolution in CRT screens don't look pixelated.

He might not be old-fashioned, but you are retarded.
Please tell me how something like this would look on a CRT:

You're not going to see the pixels, right?
I'm guessing it would look like this, right?

Them fucking magic CRTs making pixels disappear.

this is how they are supposed to look on actual crt, and those are closeups,

Again the same clueless stupidity.

"Here's how they are supposed to look.... on close-ups"
They're supposed to look like nobody ever seen them.

Again, funny how the Codex forum software shows how stupid you are. When your images are scaled down (as in, how people would actually see them), the pixelation is obvious.

 

[Luckmann]

Could...

Could someone do me the favor of running the Baldur's Gate 1 portraits through this shit?

 

[ortucis]

Pretty sure pixelated visuals look pixealated on CRT's.

Source: Me, who has been gaming since DOS. Games like Paratrooper, Prince of Persia and Dangerous Dave, all had pixelated look.