[MrFlibble]

Nice job on the screens themselves but the lower part of the screen case got blurred and lost detail -- I believe you were considering to use the hi-res version from old HRP files?

[Phredreeke]

Yeah, there are a ton of tiles using the same frame, that's why I asked who made those tiles for the HRP.

Here's a mockup I made using them.

Attached thumbnail(s)

• 

[Phredreeke]

Thinking about it, I never explained how I made screens in the last few upscales.

1. Upscale the image with NMKD's Typescale model (optionally preprocess it with SSAA9x)
2. Using imagemagick, shift it 2 pixels in both axises then scale to 25%
3. Apply a palette with only the intended hues (I used a copy of palettes 1,2,7,8 with the fullbrights edited out)

For some of them I edited out everything but the intended color (the yellow dots in the tile with the wireframe planet for example) and upscaled separately. Then I edited the results together.

[Phredreeke]

I decided to make an animation showing the different steps

[Phredreeke]

I decided to put the spacehatch through the same process as above. Here's the result



Is this better or worse (once the keypad is fixed) than the old frankentile using a rotated tile0447?

https://imgsli.com/MzkxMDE

[MrFlibble]

Frankentile is noticeably sharper so it's not easy to judge here I'd say.

I tried comparing the new version with a test upscale of the same tile I made a while ago by blending several layers (ThiefGold after HQ2x "softening" with something else -- it's some pages ago in this thread but I'm attaching another copy for reference), but the trick is that some parts seem to look better in one image and some in the other, and conversely, some areas have more undesirable artifacts or bad looking parts in one image and other areas on the other.

How I understand it, the only way to get a definite improvement so far is to have an artist go over the upscale and fix problematic parts. I'm afraid that invoking more AI models ultimately servers only to multiply the ranges of "relatively good" and "relatively bad" results.

[Phredreeke]

A pair of new upscales




Slide-over comparison for the former https://imgsli.com/Mzk0OTI

These use mostly the previously described Detoon-thiefgold interpolation, with sections cut in from NMKD-Typescale upscale.

[MrFlibble]

I'd say moderately nice job overall, although it feels that some downscaling artifacts are visible, e.g. the slight ringing around black outlines on the monitor and panel.

As for the computer screen content on the first image, the red shape at the bottom appears to be some kind of dynamic bar graph (like in Winamp -- no idea how to call it properly), which it totally lost in the upscale.

[Phredreeke]

An attempt to restore the dithering lost in the upscale process

[Phredreeke]

I've continued working on the various screen/control panel textures. All of these used NMKD-Typescale with pixelshifted downscaling to some extent for the displays

Spoiler

Bottom part uses the new dithering restoring method I showed in the last post. Also, several Plutonium Pak tiles share the same frame with just different screen content


Note: "Text" coloring unfinished. This and the next one has a row of garbage pixels (actually part of the top of the tile that was wrapped around) that needs to be edited out


This one was a major pain in the ass. Still want to touch up some parts of it

[jkas789]

Those screens are looking really clean. Damn that must have taken some time (you said the second one needed some manual touch ups on the discord right?). Congrats on the results of your hard work mate, IMO they look good!

[Phredreeke]

they ALL needed manual touchups (the second one was slightly more touched up compared to the version I posted on the art channel in discord, in the blue section on the middle panel)

[Phredreeke]

Last night I decided to tackle a pair of tiles already in the upscale pack

Spoiler

The old version (from my memory) was edited together from two different upscales, one being an interpolation of XBRDD and Fatality3 models, I can't remember what the other was.

While the text for FEDERAL LOAN AND TRUST still uses that interpolation, I decided to attempt a few different models for the background texture. In the end I settled for using the box model downscaled to 50% size and ran through my area-based repalettisation script



This is noisy, but with some color substitutions I got a result I was happy with. I sent it to Tea Monster, who made a textless edit of the background texture.



While I was happy with the FEDERAL LOAN AND TRUST text, the smaller text below still looked rather poor. Also, since the text is shared between those two tiles and an additional two tiles in Duke it out in DC it would be preferable the result is not legible. At first Tea Monster produced an example of gibberish characters to use. I did some tests with those, including downscaling them and upscaling again to "degrade them"

then I tried a different approach. I grabbed the text from the old upscale, placed it over the textless version then resized it down to 90%, I then moved the individual "words" to align roughly with that of the original. Now we got the same text but with more space inside it.

As a final touch up, I made a new layer of the image with a threshold effect set (making everything below threshold black and everything at or above it white), applied a convolution filter subtracting the pixel above and to the left, inverted and softened it, then set it to multiply blend. I had not originally intended to process the bolts or the landmark sign on the bottom of the sign like that, but I liked the result so I kept those in as well.

For the San Andreas Fault I decided to touch up parts of it by 1. doing a separate upscale where half the letters are omitted, for example see the side of E 2. using parts of letters from the Federal Loan and Trust sign, as that came out cleaner

Slide-over comparison https://imgsli.com/NDA2OTQ


Also, first DC upscales

[darkprince227]

Looks like progress is being made. Not to sound rude but maybe make the smaller text actually readable?

[Phredreeke]

I would have to come up with a text that makes sense for both tiles. Hence why I kept it illegible.

The "Missiles" label on the devastator ammo has been yeeted


As with the small text on the two tiles on my last post, the "text" here is still based on that of the original tile although the way I went about creating it was different (I isolated just the grey part of the label and upscaled with Fatal Text model, this gave a good shape but blew out any texture on it. So I took a copy of just the "text" part, duplicated and resized a bunch of times. Then I placed the upscaled text shape over that with blend mode set to multiply)

BTW I boomered and linked the wrong slide-over comparison above. This is the correct one https://imgsli.com/NDA3ODE

[darkprince227]

 Phredreeke, on 14 February 2021 - 07:07 AM, said:

I would have to come up with a text that makes sense for both tiles. Hence why I kept it illegible.

The "Missiles" label on the devastator ammo has been yeeted


As with the small text on the two tiles on my last post, the "text" here is still based on that of the original tile although the way I went about creating it was different (I isolated just the grey part of the label and upscaled with Fatal Text model, this gave a good shape but blew out any texture on it. So I took a copy of just the "text" part, duplicated and resized a bunch of times. Then I placed the upscaled text shape over that with blend mode set to multiply)

BTW I boomered and linked the wrong slide-over comparison above. This is the correct one https://imgsli.com/NDA3ODE

What does the HRP texture on it say? Maybe use that as a basis?

[Mark]

the HRP has the correct text on each one

[MrFlibble]

The text kinda looks like Tengwar

Also apologies for my ignorance, but what is the round logo at the bottom of each plaque supposed to represent?

[Phredreeke]

I believe it's a bear which is the state animal of California

[MrFlibble]

I also thought that it looks like a bear but wasn't sure for the very small size of the shape in the original image (could as well be a depiction of a mountain or something).

BTW forgot to tell that I like your new computer panels.

[Phredreeke]

Speaking of computer panels... Here are a few I did a couple of days ago. After some attempts at upscaling ended up with the keyboard being lumpy, grainy and indistinct (Tea Monster's words) I used the HRP version as a basis for it

[MrFlibble]

This is an excellent illustration of my current views on the whole AI upscale thing. At least when we're talking about images with lots of small details like the panels above, an artist's take on the high-res version is way ahead of the best AI upscales available. Even though you can see some differences from the source material, especially on the first beige panel, I think it's still preferable compared to an upscale, even the latter is techincally more accurate to the original low-res image.

I also believe that a little change/novelty here and there in texture detail will not hurt or compromise the HRP. After all, many official high-res releases (e.g. the console/Mac textures from Wolfenstein 3-D do the same).

[Phredreeke]

A tile lifted from LameDuke then recolorised for the final game. The final game's tile to the left for comparison



A couple of tiles me and Tea Monster did in the last couple of days

(the former was already in the upscale pack but the new version is cleaner)


Finally, this used Dedither model followed by Nickelback, then run through my dither restoring script and repalettisation script. Slide-over comparison: https://imgsli.com/NDE1MDg

[MrFlibble]

 Phredreeke, on 20 February 2021 - 05:09 AM, said:

A couple of tiles me and Tea Monster did in the last couple of days

You two made quite a fabulous dream team

 Phredreeke, on 20 February 2021 - 05:09 AM, said:

Finally, this used Dedither model followed by Nickelback, then run through my dither restoring script and repalettisation script. Slide-over comparison: https://imgsli.com/NDE1MDg

I'm not sure what this is supposed to represent but since this is apparently some kind of artificial metallic structure, I guess some of the straight lines (like the black hexagonal shape in the centre) could look less jittery. I understand that the original low-res image likely tries to convey the idea that the metal is worn or dented in places, but the upscale kinda looks like the image was redrawn with an ink pen or something. But maybe it looks better in-game.

[Phredreeke]

...and this is what my palettisation script was meant to avoid lol



This is caused by greys being interchanged with beige colors, which with pal02 end up much brighter. With there being beige present in the original tile makes it so that any pixel in the zone that pixel belongs to can be mapped to beige.

[MrFlibble]

Speaking of palettisation, I have to say I've got a bit of an odd feeling on this. It does dithering just fine, but the resulting images have this look that kind of resembles a high-clour photograph that was converted to a GIF image. Or as if a level of noise was deliberately added to the texture. The low res originals also have dithering but it looks, I don't know, more natural. Maybe because it was likely done by hand.

[Phredreeke]

The redithering is separate from the palettisation script.

Also I noticed that my redithering script ihad swapped the two inputs, causing it to subtract the original from the dedithered image instead of the other way around. with uniform dither either pattern will look similar, however when there's very small details being deleted those are inverted. I've also made another script that will attenuate the dithering when it's particularly strong.

The redithering script will only used when there's dithering in the original tile.


Edit: The fix for the above tile. Make sure to check out the different options on the menus on top of the image https://imgsli.com/NDE2Njg

[MrFlibble]

Honestly I have not noticed much difference between upscale and upscale2.

I think it's fine overall, but I think the black outline on the top of the central embossment (marked with red rectangles below) was broken up a bit too much by dithering (compared to the original low-res image), it should be not too problematic to fix that.

As for the green circles, these are areas which I call ringing that are apparently an after-effect of sharpening. Essentially there is a greater contrast between the light and dark areas than in the original low-res image, which has produced a darker outline of especially bright parts -- most noticeable on the left side of the image where there is a "shine" on the edge of the oval embossment. You can see in the upscale a thin dark line that "shadows" the shiny edge directly below it. I'm not sure if you can do anything with this but I point this out as a recurring artifact in your upscales that I think is not really desirable -- unless there is no way to avoid it without losing quality of the result.

This kind of ringing is clearly not produced by ESRGAN as using a different model does not generate it (this is 2xFakeFaith -- not very good looking but simply an example):

[Phredreeke]

there's not a big difference between upscale and upscale2 with the default palette.

now if you switch it to pal02 there's a pretty big difference: https://imgsli.com/NDE2Njg/2/3

[Phredreeke]

Haha gun goes pew pew

https://imgsli.com/NDIzMzE
https://imgsli.com/NDIzMzA

More first person sprites