[Phredreeke]

https://imgsli.com/MjQzODA
https://imgsli.com/MjQzODE
https://imgsli.com/MjQzODI

Edit: Lollypop has some weird side effects (scaled 2x afterwards for clarity)

[MrFlibble]

 Phredreeke, on 04 October 2020 - 01:01 PM, said:

Edit: Lollypop has some weird side effects (scaled 2x afterwards for clarity)

Honestly I think dev portraits should not be upscaled. Hunt down the original pics, or use any other pics of them if you need, or don't touch the low res stuff at all.

Anyways, I interpolated Lollypop with Rebout Blend at 0.5 which seems to produce clean results that seem rather accurate to the low-res originals.

Spoiler

ERMAHGERD

BERKS

 Phredreeke, on 04 October 2020 - 09:20 AM, said:

Is mtpaint able to batch process?

Not that I know of. But the Dithered effect is rather similar (albeit, I believe, better) to a method used by GIMP.

 Jimmy, on 04 October 2020 - 10:51 AM, said:

I'd also recommend using Image Analzyer to convert the images back into palette. I haven't used mtpaint myself, but I've seriously never found another program that even comes close to Image Analyzer, even ones that cost hundreds of dollars. It appears most programs just grab the closest color it finds first, whether than the best color match it can find which Image Analyzer does.

I got Image Analyzer a while ago when you mentioned it in a similar context. It is indeed very good at finding colour matches when palettising (certainly better than other tools that I tried), except I have found no way to use dithering with a specified palette.

[Mark]

The higher res developer head pics are posted in these forums somewhere. IIRC the heads needed to be added to the original backround which could be rescaled now. I made one just to start the ball rolling for anyone else to continue with. If you can't find the thread with the pics I'll search for them on my backup drives. I know I saved them somewhere in case I decided to remake all of them.

edit: I started searching here and found a post of mine from Jan 2018 that also mentioned someone found the highres heads and that I remade 1 tile. So concentrate your searching before that date.

[Jimmy]

LollypopWhatever Mr Flibble is doing looks awful.

[Phredreeke]

It's not Lollypop doing that, it's the high amount of Rebout being interpolated with it. Some pixel patterns (the buttons on the space door panels) gets blurred out of existence in pure Lollypop, I assume MrFlibble was trying to mitigate that.

[Phredreeke]

Alright, I've upscaled the tiles using lollypop interpolated with manga (not very time consuming), then tried making it play nice with palswaps (VERY time consuming)

https://imgsli.com/MjQ1Mjc
https://imgsli.com/MjQ1Mjg
https://imgsli.com/MjQ1Mjk
https://imgsli.com/MjQ1MzA
https://imgsli.com/MjQ1MzE
https://imgsli.com/MjQ1MzI

First five compares it with the previous tiles I processed. Last compares the palswap with and without processing for compatibility.

Also, I made these the other day, after improving the palswap compatibility of my weapon upscales.

Spoiler

[jkas789]

 Phredreeke, on 05 October 2020 - 04:55 PM, said:

Also, I made these the other day, after improving the palswap compatibility of my weapon upscales.

Spoiler

Quick question dude. Are this the "indexed (is that the terminology?) tiles" that you mentioned you were working on in the duke4 discord some time ago.

Also from my untrained and user end eyes, the manga/deviant pixel looks blurrier (is that the correct word, it looks smoother than the left image but it seems to loose some detail. Am I an idiot?) than the lollypop/manga. IMO.

[Phredreeke]

Yes, it's blurrier. That part is DeviantPixel (which I interpolated with as it reduces the noise inherent in the manga model, and it's less artifacty than when I interpolated with PSNR before), and also an antialias script I had used before upscaling.

For my 2x scales I added a sharpening filter while downscaling. I didn't want to risk oversharpening and also worried it would interfere with my attempts at making it palette compatible.

Also it's not as simple as one being definitively sharper, look at the fifth one, (actually I think overall it's not as much sharpness as higher contrast)

[jkas789]

Oh yeah I see what you mean there. On the fifth image the square grid is lost on lollypop/manga right?

Man the amount of effort you have to put on this must be rough, especially if you want to remain loyal to the original tiles.

[NightFright]

So I have closely checked Phredreeke's 2x/4x texture res comparison screenshots from Oct 4 and found that the difference isn't that big. There's surely a bit less aliasing on fonts as seen e.g. on the cinema sign, and textures may have a VERY small amount of additional crispness, but I wonder if that's really noticable during actual gameplay (especially if you don't have the comparison).

[Phredreeke]

Oh, it will be noticeable as I've not done any processing to improve palette swap compatibility with the 2x versions...

Once that is done however... but that's not something I'm interested in pursuing until I got a definite answer over whether we are going for 2x or 4x


Edit: one upside of using 2x scale that's not been mentioned is that it would allow the use of my existing upscales of switches and posters without further scaling

[MrFlibble]

Phredreeke, on 05 October 2020 - 04:55 PM, said:

Alright, I've upscaled the tiles using lollypop interpolated with manga

I think the interpolation with Manga does not produce better results than what you had.

I ran a few more tests and identified another model called Deviance (60k) that produces interesting results. Here's the interpolation of it with Lollypop (again, alpha = 0.5):


The Upscale Wiki page lists two Deviance models but both links lead to the same V1 download.

Jimmy, on 05 October 2020 - 11:03 AM, said:

LollypopWhatever Mr Flibble is doing looks awful.

Thank you for your kind encouragement!

[Phredreeke]

Phredreeke, on 06 October 2020 - 02:48 AM, said:

Once that is done however... but that's not something I'm interested in pursuing until I got a definite answer over whether we are going for 2x or 4x

Attached File(s)

•  mangadownscale2x.zip (9.66MB)
  Number of downloads: 205

[MrFlibble]

Those look good!

However I have to say that there's one thing in common with all the 4x models I've tried out so far, and that is they generally don't produce the pixel art look at all, which is especially noticeable at their default resolution, and simply scaling down to 2x, even with nearest neighbour as hidfan did with the Doom upscales does not make a lot of improvement (I'm not talking about the above results, rather about the tendency in general).

I just tried a 2x model called Faithful which was trained specifically to make higher resolution textures while staying true to the source material. I think some results are pretty decent (all palettised and scaled up to 4x with nearest neighbour):




It makes the colours somewhat darker though.

More examples here:
https://imgur.com/a/aP4A291

[Jimmy]

MrFlibble, on 06 October 2020 - 03:05 AM, said:

Thank you for your kind encouragement!

Not trying to put you down specifically. It's just that the methods you were using had much more erroneous behaviours and blurriness than what Phredreeke had already been doing.

MrFlibble, on 06 October 2020 - 09:22 AM, said:

I just tried a 2x model called Faithful which was trained specifically to make higher resolution textures while staying true to the source material. I think some results are pretty decent (all palettised and scaled up to 4x with nearest neighbour):

I quite like these but it's got some weird artifacting going on, almost like a JPG. I don't know what that's all about. It might be worth tinkering with this model more.

[Phredreeke]

Jimmy, on 06 October 2020 - 10:33 AM, said:

I quite like these but it's got some weird artifacting going on, almost like a JPG. I don't know what that's all about. It might be worth tinkering with this model more.

The problem with 2x models is that they can only be interpolated with other 2x models, of which there are very few compared to 1x and 4x.

[Jimmy]

The artifacting might be more noticeable because of the 4x upscale there, yeah.

[MrFlibble]

Faithful indeed produces rather faint JPEG-like artifacts, like quite a few other models do. And indeed the problem is that there isn't much room for improvement for the scarcity of other 2x models as Phredreeke pointed out. It's still much better than I initially expected though.

Jimmy, on 06 October 2020 - 10:33 AM, said:

Not trying to put you down specifically. It's just that the methods you were using had much more erroneous behaviours and blurriness than what Phredreeke had already been doing.

I try out different models available and pick those that produce at least moderately acceptable results (many result in unusable garbage). My idea here is that while there seems to be an optimal model for now (at least when ESRGAN is concerned) it is well worth it to be on a lookout for potentially useful new models, especially considering how fast they develop and improve.

I genuinely don't understand Phredreeke's predilection for the original Manga model. It is pretty old and while it is good at some things, it also appears to be responsible for certain quirks that could be improved upon. For example, some lines that are definitely meant to be straight become bent or distorted.

Like here Manga109Attempt mangles simple rectangular shapes while Rebout Blend at least keeps them more or less the same, and for Faithful this is not even an issue:


If I'm pointing out things like this out it's only because I want the HRP textures to be made in the best possible way, and look as authentic as possible. It is not my intention to undermine others' work or something. It's just that neural upscales are a continuously developing field and sometimes it might be a good idea to try for new stuff even if the currently established methods produce good results.

[Phredreeke]

When downscaled to 2x scale, those artifacts are less of an issue.

I don't think there's one model that works for all textures. A significant number of textures will need manual edits.

I've processed the lollypop-manga upscales similar to the previous ones (when I put the set together, I seem to have overlooked a few that were in the previous one however...)
I'm also attaching a set of manual edits I made prior for the upscale pack. Most of those lack the sort of palette compatibility processing in the ones I recently did.

Some comparison between those edits to the tiles in the previous (manga) pack in the spoiler tag

Spoiler

Attached File(s)

•  upscalepackwiptiles.zip (687.61K)
  Number of downloads: 165
•  lollypop-manga-downscale.zip (9.7MB)
  Number of downloads: 162

[NightFright]

Text on textures is difficult in general. Letters either tend to "bleed out" or get aliased. What needs to be avoided is that textures actually lose detail or become so blurry that it looks as if bilinear filtering has been applied.

I guess supersampling (e.g. upscaling to 4x or higher and then downscaling to 2x) is also not really an option if the algorithm actually loses details as resolution increases.

[Phredreeke]

Yes, for certain tiles I cut in from an upscale using another model, or sometimes I overlaid letters from other sources (the e in Phone)

Also, for the tiles in the set from the upscale pack, more often than not sharpening was applied as it was downscaled. I can see about making a sharper downscale of the manga-deviant set (as in most cases, it ended up the better looking relative to the lollypop-manga set) as I originally avoided sharpening for reasons detailed earlier in the thread

[MrFlibble]

Phredreeke, on 07 October 2020 - 07:25 AM, said:

When downscaled to 2x scale, those artifacts are less of an issue.

I wonder if there is a method of downscaling to 2x from 4x that would add a more pixel-y look to the result? From my observations, palettising and even applying dithering alone does not really help much. Most images made from 4x upscales still feel like they were repainted by an artist with a completely different art style compared to the original.

I have tried out various downsampling methods available in image editing programmes, but they all are more or less the same in that the resulting image has this stamp of being downsampled on it. It might be not always noticeable but it's still there. The problem is, even if you palettise first and then scale down with nearest neighbour it does not make much of an improvement.

I get it that the issue is rather complex because the original low resolution definitely dictated both the amount of detail that could be put into a texture, and the ways in which these details could be expressed.

Phredreeke, on 07 October 2020 - 07:25 AM, said:

I don't think there's one model that works for all textures. A significant number of textures will need manual edits.

I completely agree with that but I'm still curious. What advantages in Manga make it a viable choice in your opinion?

[Phredreeke]

It's a trade-off.

[Jimmy]

MrFlibble, on 07 October 2020 - 05:28 AM, said:

Faithful indeed produces rather faint JPEG-like artifacts, like quite a few other models do. And indeed the problem is that there isn't much room for improvement for the scarcity of other 2x models as Phredreeke pointed out. It's still much better than I initially expected though.

....RAMBLING....

If I'm pointing out things like this out it's only because I want the HRP textures to be made in the best possible way, and look as authentic as possible. It is not my intention to undermine others' work or something. It's just that neural upscales are a continuously developing field and sometimes it might be a good idea to try for new stuff even if the currently established methods produce good results.

Me criticising your work doesn't negate any of this. I've criticised Phredreeke's work more than you, but he's also been doing it longer. It's not a personal attack. What you came up with was pretty ugly. Phredreeke's method produces erroneous behaviour pretty often that's true, but it is still the best method yet to be devised, and he puts in extra work to avoid those issues or work around them. That's all. I have the same goals here as you do, and that's why I mention when something doesn't look good.

[Phredreeke]

and here's the sharper manga downscale I talked about before, along with slide-over comparisons

https://imgsli.com/MjQ3MjQ
https://imgsli.com/MjQ3MjY
https://imgsli.com/MjQ3MzI

Attached File(s)

•  7manga-sharp-pal.zip (10.93MB)
  Number of downloads: 162

[MrFlibble]

Yesterday I found a model not listed in the Upscale Wiki called Clean Pixels (get it here). Initially it did not impress me a lot but I ran a few tests and I think it's rather decent.

This time I downsampled to 2x by first applying the pixelise filter in GIMP, then scaling down with nearest neighbour. Palettising done with ordered dithering (although I'm not sure if it looks good on everything).

[Jimmy]

Now that looks nice.

[NightFright]

That poster babe without eyes gives me the creeps, though.

[Phredreeke]

Here are some I think MrFlibble will like

Spoiler

They use Lollypop interpolated with LyonHrt's XBRDD model (at 0.4)

[Mark]

The ones a couple posts up look good except for the distorted blue ring. I noticed it in other posts too.