[Phredreeke]

One important aspect. Scalertest (to my knowledge) does not let you set scalers other than xBRZ from the command line. I decided however to do some tests using xBRZ with 3x scale factor. This was downscaled using triangle filtering with imagemagick and then upscaled again using an interpolation of detoon and thiefgold

https://imgsli.com/MzY2MDk
https://imgsli.com/MzY2MTA

The second of these two is a significant improvement over using the dedither model.

This merits further testing I think

[MrFlibble]

If you can counter the blurriness on the XBRz upscales then this would probably be a more preferred method to dedithering.

One thing though, unless you have tested that already, it's probably worth checking if 3BRz is indeed preferable to 2BRz or 4BRz, because I'm not certain that they would behave the same as HQx preprocessed images in this respect. Just in case you know.

Also I never used triangle filtering. Does it give a specific advantage? Again, I'd suggest trying out different downscale methods and see how they affect the result (some might produce less blur, possibly).

[Phredreeke]

https://imgsli.com/MzY2OTE

Did the same post-processing on both (downscale then run through the unresize model, then use a custom script that mixes in color from manga upscale)


and here it is downscaled with imagemagick's -scale instead of triangle (which use an algorithm which is called pixel mixing)

https://imgsli.com/MzY2OTY

It's clearly sharper, but some of the artifacting from dedither makes a return

https://imgsli.com/MzY2OTc


One thing I had neglected to test was using Thiefgold without interpolation. Turns out, uninterpolated ThiefGold with triangle downscaling gives almost as sharp results as Detoon-Thiefgold interpolated using pixel mixing downscaling, but with less artifacts

https://imgsli.com/MzY3MDA

[Mark]

Are you guys still in the experimenting stage with all of these different models and programs or has a pack of acceptable textures been voted on and is being expanded as you progress? With dozens of options to choose from for every texture it seems like its going to take a long time to complete and have coherence. I don't envy the job ahead of you.

[Phredreeke]

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

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

No one stepped up.

I like Dedither + Detoon/thiefgold for most textures.

In addition to the models and preprocessing used, there's also the matter of applying palettes so that they look nice with palswaps

Edit: I uploaded a set of comparison images. This is just original + Dedither/Detoon-ThiefGold combo in two versions, with and without adaption for improved palette compatibility. However, when the original palswap is noisy enough, the version not adapted for palette compatibility looks better. Below is an example

Attached File(s)

•  forcomparison.zip (7.32MB)
  Number of downloads: 204

[DNSkill]

Phredreeke, on 06 January 2021 - 01:53 PM, said:

I uploaded a set of comparison images.

Not sure if this is the type of help you were wanting, but I tried:

This is in response to this file attachment in your most recent post on this thread.

In my observation, I find the third options, the ones on the right-side, are the best variants for every tile included. There's a few things I wanted to bring to your attention that I saw that might be worth considering an extra look at:

• tile0150 - primarily the control panel portion with the keypad on it. It's hard to notice the keys in the upscale, and the red tinted variant could use some work in this same area to look more like the original tile. The rest of the door is fine, though.
• tile0153 - KEEP CLEAR is kind of hard to read, could use some work to make it easier to see these words. The rest of this tile looks spot-on otherwise.
  
• tile0227 - I think this could have a little more shading to the little crevices. They seem rather faint in comparison to the original. This may just be a nitpick, though, but I thought I'd mention it anyways.

So, other than that, I would say the right-sided tiles are the best, and those 3 mentioned above would be the only ones I'd suggest making a few further edits to.

[Phredreeke]

The red version is the same as the one above - I can't edit one without changing the other. The intent is for the engine itself to generate the palswaps.

I will attempt using another model for 150 and 227 tomorrow, and I'll send Tea Monster 153 to see if he can replace the text.

[Phredreeke]

Tile 150 before and after touching up the keypanel

[Tea Monster]

Modern problems require modern solutions.

Still WIP, but looks promising.

[darkprince227]

It's all looking cool so far

Does anything think we'll be seeing Version 1 or download of this new pack anytime soon? (I forgot what it was called)

Gods know we need something to look forward to this year

Whatever the case, keep up the good work everybody

[Phredreeke]

Sharing a couple of upscales

Spoiler

(dedither + detoon/sourcetex2 interpolation upscale, 50% downscale, unresize, custom palettisation script)

(50% downscale, Nickelback upscale, custom palettisation script)

(dedither, fatal photos upscale, 50% downscale, FFT denoise, gamma adjusted noise, palettisation)

(palette swap (pal17->pal13), dedither, fatal photos upscale, color copied back in, 50% downscale, custom palettisation script)

Image 1, 2 and 4 use a new palettisation script which divides both original image and upscale into smaller tiles, and palettises the upscale using the colors in corresponding tile in the original image.

[MrFlibble]

Tea Monster, on 14 January 2021 - 04:09 PM, said:

Modern problems require modern solutions.

Still WIP, but looks promising.

Awesome work on the pistol model, but I can't but point out the stark contrast between the crisp detail on it and the blurry upscale of the trooper himself. I think it would not become much more noticeable if you put that pistol on the original, low-res sprite.

I guess the idea is to sharpen the base sprite and blur the pistol to get them to some kind of common denominator, but at this point I think it becomes obvious that sprites made from high-resolution models would look better than upscales, no matter how much effort is put into refining them.

I know it's a lot to ask, but what do you think of making a full model that would be faithful to the low-res original and then using it to produce high-res sprites? Possibly leaving ESRGAN upscales for textures only? Or at least leaving the upscaled sprites to be gradually replaced?

[Tea Monster]

Cheers. It's still WIP, so we are playing with lighting and how much blur to apply to get it looking right. Not at all finished yet. Lots of farting about to be done.

The modelling idea is great and probably the best way to get ultra-sharp sprites that look like the originals.

I've just finished making a ton of models for the HHR mod and they weren't well received. I don't really feel like spending another two years or so making all the enemies (yet again). I might finish my Cycloid model one of these days. He's done to look like Chuck Jones' model.

[Phredreeke]

Keep in mind the sprite in Tea Monster's screenshot is of the upscale before it is downscaled to the final size. This is how it looks after downscaling/sharpening/palettisation (without the new gun)



(I need to undo the eyes for the "ERP" version)

Also, I really don't want to burn out Tea Monster.

[Phredreeke]

An attempt I made at tiles 0505-0509



anyone know where this one is used in game?

[Jimmy]

It's the animation for glass maskwalls breaking.

[MrFlibble]

Phredreeke, on 18 January 2021 - 08:52 AM, said:

Keep in mind the sprite in Tea Monster's screenshot is of the upscale before it is downscaled to the final size. This is how it looks after downscaling/sharpening/palettisation (without the new gun)

Unfortunately downscaling/palettising does not entirely solve all the issues with AI upscales that we have.

Don't get me wrong, your result is very good. But with what ESRGAN can do at this moment, whatever resolution the upscale ends up at is still low on details compared to a hypothetical image that was created at the target resolution from the start (e.g. from a hi-res model). I think this might partially account for the slightly "blurry" look of the resulting sprites.

ESRGAN sharpens/brings out the details that are present in the source image, but can't add more than that. That was my point. The trick here is that when the source has enough detail the upscale might look "bland" but since the image is complete (e.g. you see the complete shapes of the Trooper's head, torso, limbs etc.), it is not as noticeable. But with the pistol, it was reduced to a handful of pixels in the original sprite so ESRGAN has trouble bringing up a coherent shape, hence it becomes apparent that the upscale is imperfect. But this imperfection is implicitly present in the entire image, not just where it stands out. But that's how I understand the entire problem, currently.

[Phredreeke]

Jimmy, on 20 January 2021 - 11:28 PM, said:

It's the animation for glass maskwalls breaking.

I'm a fucking idiot. that's the tile preceding those... I thought for some reason it was some kind of screen static

Edit: Did another attempt at it


Edit 2: Recorded it in game. This is slowed down to 20% speed

[Jimmy]

It's okay. It's something you only see for a split second at any time. Most people probably don't even know there's an animation there.

[Phredreeke]

I had initially only intended to use this model for the glyphs, but Tea Monster liked the result so much he suggested I'd use it for the whole display






(these have transparency, but it may be tough to tell given the forum's background)

[MrFlibble]

Why did the circles come out so jaggy in the second pic?

And the wireframe globe in the third image, I guess it was supposed to have regular shaped sectors/cells but the upscale messed them all up.

All these are very simple shapes though, I think it would make sense to just remake them in high-res manually if the time permits. No need for AI if the result does not really compare favourably even to xBRZ style algorithms.

[Phredreeke]

Because the original tile wasn't antialiased (because I only intended to use them for the alien glyphs)

I upscaled the 3rd tile again with ssantialias9x pre-applied. I prefer this look tbh

[Mark]

I'm assuming the planet grid looks like it was made from spaghetti noodles because the original sprite is poor quality.

[Phredreeke]

Original tile



Perhaps the wireframe planet would be better recreated, as long as care is taken to align the grid size and position to match the original.

Here's another one I just made

[Mark]

Teamonster could whip up a globe in Blender. Make the lines a little thicker and make the backround transparent in photoshop. It shouldn't take long at all. The 2nd pic is a quick attempt with transparent backround but not in the duke palette.

Attached thumbnail(s)

• 
• 

[Phredreeke]

Potential options (Tea Monster helped render and place the sphere, as my attemps to failed pretty badly)

[Tea Monster]

Nonsense, you've been a trooper on this project.

[Phredreeke]

Added in the yellow dots and here's the result

Edit: Animations

Attached thumbnail(s)

• 
• 
• 

[darkprince227]

Was it always animated, or am I imagining things?

[Phredreeke]

Upscales of a couple of displays