[Phredreeke]

I've experimented more with ThiefGold, more specifically I interpolated it with DeToon which reduced some artifacting in the model.

Here's a tile I worked on yesterday


Tea Monster pointed out the texture on the bricks made them look like wood. I got the idea to replace the bricks with grey before upscaling


Afterwards I tweaked the levels and recoloured the bricks brown again. Here's the downscaled result

[MrFlibble]

 Phredreeke, on 01 December 2020 - 07:46 AM, said:

Edit: I attempted processing the tile in question with the ThiefGold model - it gives a slightly better result in regard to the top piece

Both CleanPixels and Lollypop do a decent job with the top piece:


Fatality MKII produced a really creepy effect on the face and the rest is not super impressive so I'm not showing that one here.

The Thief Gold model made the whole image look like it's covered with rust -- not surprising if the model was trained for Thief Gold textures?

Also both CleanPixels and Lollypop also work nicely with the lines and edges on the Duke Burger sign:


The way Lollipop interpreted dithering is really weird though.

And Charsprite is the most primitive but also the closest to the original image (in a way):

[Mark]

All 3 of those would require a lot of manual touchup. I noticed all 3 produce the same error at the bottom of the D.
The inner red D trim is obviously raised above the yellow behind it. The outer D looks sunken in instead out front like the inner one.
The uneven edges seems to be a familiar theme with upscales too.This might be one of those better created from scratch. What do you think?

[Phredreeke]

Not to mention, Tea Monster already touched it up...

Edit: Here's another tile upscaled with DeToon/ThiefGold interpolation. Left is raw upscale, right is with postprocessing.

[MrFlibble]

Mark, on 03 December 2020 - 10:08 AM, said:

All 3 of those would require a lot of manual touchup.

I was specifically talking about the rectangular shapes like the box with the burger counter, not the entire image (see Tea Monster's post above for the manually edited tile).

Phredreeke, on 03 December 2020 - 10:33 AM, said:

Edit: Here's another tile upscaled with DeToon/ThiefGold interpolation. Left is raw upscale, right is with postprocessing.
detoonthiefgold-0789.png

If the rest of the textures comes out as good with this one it might be the best option ever.

Do you know what this Thief Gold model was trained on? Who's the author?

[Phredreeke]

MrFlibble, on 03 December 2020 - 12:59 PM, said:

I was specifically talking about the rectangular shapes like the box with the burger counter, not the entire image (see Tea Monster's post above for the manually edited tile).


If the rest of the textures comes out as good with this one it might be the best option ever.

Do you know what this Thief Gold model was trained on? Who's the author?

Spoiler

Akven 2020-05-15
@news @wiki
Name: 4x_ThiefGold_110000
License: No idea
Link: https://drive.google...iew?usp=sharing
Model Architecture: ESRGAN
Scale: 4
Purpose: Various game textures. Primary wood, metal, stone

Iterations: 110000
batch_size: 8
HR_size: 128
Epoch: ?
Dataset: Thief Gold & Thief II community made textures
Dataset_size: don't remember
OTF Training No
Pretrained_Model_G: RRDB_ESRGAN_x4

Description: This model was based on community made HD textures for Thief Gold & Thief II: The Metal Age. Textures were reduced to 256 colors to simulate Thief Gold palette. Surprisingly enough, model works good with any kind of textures with the exception of brights colors which sometimes generate dotted artifacts. Especially good it upscales stone, metal and wood surfaces. It is a good model to upscale game textures especially from old games with low quality sources. Working of Thief Gold/Thief II ESRGAN mod I used this model 50% of the time. Often it gaves better results than Skyrim trained models from wiki. Example: https://imgsli.com/MTYzNjI (not sure from ThiefGold or ThiefGoldMod model).

Name: 4x_ThiefGoldMod_100000
License: No idea
Link: https://drive.google...iew?usp=sharing
Model Architecture: ESRGAN
Scale: 4
Purpose: Various game textures. Primary wood, metal, stone

Iterations: 100000
batch_size: 8
HR_size: 128
Epoch: ?
Dataset: Thief Gold & Thief II community made textures
Dataset_size: don't remember
OTF Training No
Pretrained_Model_G: 4x_Manga109Attempt

Description: Version of the previous model but based on Manga109 pretrained model and with slightly different dataset. Sometimes gives better results especially for wood and metal, sometimes worse. Sometime generates the same dotted artifacts on very bright/white images. Still better most of the time for Thief game series than Skyrim based models.

[Tea Monster]

Looking at some of the HRP tiles, some of them are truly woeful attempts. Phredreeke suggested that it might be a good basis for HRP v2 if we use some of the tiles from this effort over there.

[Phredreeke]

To clarify, I was talking specifically about the developer faces I did for Faces of Death and Pigsty, since they're not something you could simply remake.

[Mark]

Tea Monster, on 04 December 2020 - 01:47 AM, said:

Looking at some of the HRP tiles, some of them are truly woeful attempts. Phredreeke suggested that it might be a good basis for HRP v2 if we use some of the tiles from this effort over there.

What tile numbers are some of the worst ones?

[MrFlibble]

Thanks Phredrekee!!!

Those Thief models are simply amazing Apologies for being sceptical earlier -- the results with textures are miles above what I've tried before.

Here's some raw upscales of random Duke3D wall textures with 4x_ThiefGold_110000, compared against 4x_CleanPixels:
https://imgsli.com/MzE5NDU
https://imgsli.com/MzE5NDY
https://imgsli.com/MzE5NDc
https://imgsli.com/MzE5NDk
https://imgsli.com/MzE5NjI

For these, upscales with ThiefGoldMod looked better to me than ThiefGold:
https://imgsli.com/MzE5NDg
https://imgsli.com/MzE5NTM
https://imgsli.com/MzE5NjE

Many of the ESRGAN models (but not CleanPixels shown above) add some sort of noise to the images to imitate detail, but only these ThiefGold ones make the end result look mostly like a meaningful high-resolution image instead of random rubbish (or in some cases, like a painting but that's not useful for video game upscales).

However, I still believe it's nice to keep some other models in mind. For example, here's the wood crate texture by 4x_Lady0101_314000_G (left) and by 4x_ThiefGoldMod_100000:


I'm thinking that maybe in the above case, Lady does a better job with the texture. And possibly here too:

[Phredreeke]

one thing you may want to try is running the texture through a dedithering model (it has a smoothing effect, even if the texture has no perceived dither in it) before thiefgold - it helps in some cases where thiefgold on its own is too noisy (see the battlelord door for example)

[Jimmy]

Thief Gold appears to favor over-sharpening.

[Phredreeke]

Here's a comparison of tile 1102 upscaled with and without preprocessing with alsa's dedither model

https://imgsli.com/MzE5NzA

Edit: Also an example showing the artifacting some images causes ThiefGold, and how it's reduced by interpolating with a different model

https://imgsli.com/MzE5NzE

[Jimmy]

I know you're just illustrating a point, but that map shows the larger San Francisco area, for future reference.

[Tea Monster]

Lady looks very painterly, but it looks a lot less artifacted and a ton more natural.

[Phredreeke]

Another trick if a model gives you problems with a particular image: rotate the image 90 degrees before upscaling it!

https://imgsli.com/MzIwNDQ

[Tea Monster]

Phredreeke, on 04 December 2020 - 05:04 PM, said:

Here's a comparison of tile 1102 upscaled with and without preprocessing with alsa's dedither model

https://imgsli.com/MzE5NzA

Edit: Also an example showing the artifacting some images causes ThiefGold, and how it's reduced by interpolating with a different model

https://imgsli.com/MzE5NzE

LOL, what would you get if you hard-upscaled the tiles and ran just the de-dither on them? We would have killed for that tool in and of itself when I was working on newspapers in the 90s. The amount of shit people would send in as camera-ready artwork that were newspaper clippings, or (barf) faxes.
That de-dither tool is amazing.

Mark: I didn't see your post earlier. The Guilty and innocent signs (1210 and 1211) are a prime example. The HRP tiles that are supplied (not my models!) have completely the wrong type of font. On both of them, they mix serif for non-serif fonts. Bleh!

[Phredreeke]

Also this one

[MrFlibble]

Phredreeke, on 04 December 2020 - 02:09 PM, said:

one thing you may want to try is running the texture through a dedithering model (it has a smoothing effect, even if the texture has no perceived dither in it) before thiefgold - it helps in some cases where thiefgold on its own is too noisy (see the battlelord door for example)

I was reluctant to try that out yesterday because my previous tests with the dedither model resulted in some detail loss in the original image (even though the results are generally good). But the Battlelord door result you've shown indeed is an improvement -- I think some lines have been straightened too compared to the raw unprocessed input image.

Phredreeke, on 05 December 2020 - 05:36 AM, said:

Another trick if a model gives you problems with a particular image: rotate the image 90 degrees before upscaling it!

https://imgsli.com/MzIwNDQ

Yeah, I noticed that too, with the second texture of the hatch with the blue keypad (which is rotated 90 degrees compared to the other hatch texture) -- it came out way worse than the one I've shown above.

[Phredreeke]

Yes, the blue keypad that you mention gets blurred out of existence by the dedither model. Fine details like that may need to be adjusted manually either way

[Phredreeke]

Decided to put up some comparisons between the latest attempts and the previous ones (new ones last)







Slide-over comparisons (new ones first)
https://imgsli.com/MzIxNjI
https://imgsli.com/MzIxNjM

[MrFlibble]

That piping is the prime example I think, it impressed me how ThiefGold got it all right when I think all the previous models that I tested failed miserably at least in some detail.

On a side note, I would appreciate it if the imgur examples were posted as separate images, as I sometimes like to open versions of an image in different browser tabs (or in different image viewer windows if locally) and switch between them, that gives a good comparison (better than side by side IMO). Can't speak for others but someone else might prefer this method too.

[Phredreeke]

Ironically I've prepared a big set of side-by-side images for evaluating whether the new process is an improvement for each given texture...

I could use a volunteer or two to help me go through them. I need them sorted into categories 1. Use old upscale 2. Use new upscale 3. Use new upscale but needs further edits

[MrFlibble]

Phredreeke, on 07 December 2020 - 03:22 PM, said:

I could use a volunteer or two to help me go through them. I need them sorted into categories 1. Use old upscale 2. Use new upscale 3. Use new upscale but needs further edits

I think it could be a good idea if the prospective artist who will be doing the edits (Tea Monster I guess?) would take a look, so that the images would be evaluated from the standpoint of someone who knows what kind of editing needs to be done and how.

[Phredreeke]

I don't want to overwork Tea Monster though, he's done a lot of work for it already.

I tweaked the process a little bit for certain tiles



First two both use ThiefGold (interpolated with Detoon) but slightly different preprocessing. Last one used Manga model.

I feel the second here looks better because it's slightly less regular than the first.

[MrFlibble]

Phredreeke, on 10 December 2020 - 06:54 PM, said:

First two both use ThiefGold (interpolated with Detoon) but slightly different preprocessing.

It looks like this slightly different preprocessing makes the model accentuate completely different parts of the image while pushing others into the background so to speak. I wouldn't say that I prefer one over the other here, perhaps the choice should be made based on what it looks in-game.

[Phredreeke]

I just realised that by shifting the image 1 pixel prior to downscaling I can get a lot crisper results

https://imgsli.com/MzMzNzA
https://imgsli.com/MzMzNzE
https://imgsli.com/MzMzNzI

(the reason is obvious once I think about it, as prior each 2x2 would correspond to 1 pixels in the original tile, while with the shift 75% of such blocks would now straddle at least two pixels from the original)

[MrFlibble]

Phredreeke, on 13 December 2020 - 08:50 AM, said:

I just realised that by shifting the image 1 pixel prior to downscaling I can get a lot crisper results

This is related to something I noticed a while ago: it seems that different image editing software handles downscaling differently in this odd 1-pixel-shift manner. Specifically, out of what I'm using, GIMP and mtPaint produce similar results, but Image Analyzer and IrfanView do the shift (or maybe vice versa). However, I've not observed a difference that would be quite like yours, perhaps it's specific to Paint Shop Pro (which I assume you're using, right?).

Anyway, out of all methods that I tried (barring straight nearest-neighbour downscale) what I liked most is (in GIMP) to first apply the pixelise filter at 2 pixel radius and then downscale with nearest neighbour.

[Phredreeke]

well, that can be attributed in part to my downscaling script doing other stuff as well

I uploaded a new gallery this time where it's just downscaling with and without shift, as well as the raw upscale for reference

https://imgsli.com/MzM1Mjg


Here's one I did today, I used the latest Spongebob model interpolated with XBRDD (at 80:20 ratio) for the middle part and downscaled with said shift. The edges used a mix of the lollypop model and manga (the latter where I flipped the image and merged using one of my blending scripts, making it more symmetrical)

https://imgsli.com/MzM1MjM


Edit: Another one I just did, using similar methods

[Mark]

It looks like upscaling is trying to outdo Linux or doom ports in sheer amount of different versions. All the different names, models, scripts, filters, etc... I wonder how long before we see just one GREAT upscaler stand out in front of the rest.