[Phredreeke]

Teamonster helped me touch up the pigcop tank

Attached thumbnail(s)

• 

[jkas789]

Oh Lard he comin' !

[Tea Monster]

[MrFlibble]

Phredreeke, on 29 October 2020 - 07:33 AM, said:

Teamonster helped me touch up the pigcop tank

Very nice!

One thing though, it seems like the Nuke logo on the frame that is the straight-behind view (second from left in the second row) is notably sharper than in the immediately preceding frame where it is viewed from an angle, presumably because the front view was pasted on from the level-end switch? Not sure if this will be as noticeable in-game though.

[Phredreeke]

Both were... maybe I should sharpen the diagonals a bit?

Ok, here's the result. Better?

[MrFlibble]

Definitely!

[Tea Monster]

[Phredreeke]

Here's my edit and downscale of the above

[Tea Monster]

[Nfsfan83]

Guys HELP:

https://forums.duke4...n-movies-style/

[MrFlibble]

The Duke Burger sign made me wonder about one thing: would the colour dithering on the brown background look different if one took a similarly shaded TrueColor image and palettised it? I'm asking because potentially ESRGAN could mess up with the source images dithering patterns and transform them into a kind of noise that is different from "authentic" dithering at the target resolution.

That said, I believe that stuff like this sign could be actually completely redrawn from scratch, not just the letters/foreground, without any detriment to the goals of the HRP, or the authenticity of the image.

[Phredreeke]

Sometimes I've tried to remake the dithering but most of the time I haven't bothered.

[Phredreeke]

I've been hesitant to share this outside of DMs because I was unsure of my capability to upscale the whole series of faces. At this point I feel more confident to show this openly.



Tea Monster did the name tag. Bobby as you can see is a composite between upscale and a vintage photo

[MrFlibble]

Looks very good! But since there's manual editing and polishing going on anyway, I think it'd be nice to clean up/straighten those outlines around the photo (I mean the light brown/black angular shape). Not sure why the pixels are so jagged in the low-res original, especially on the left-hand side, but the upscale could have them look less squiggly I think.

If you have the resources, I think many shapes involving straight lines could use some manual touch-up after upscaling and downsampling to 2x as they tend to get sorta blurry. Like the grey metallic thing at the bottom left-ish here. Is the dark blot at the centre of it supposed to be an upwards-pointing arrow BTW? The low-res version looks as if it could be but the upscale just renders it as a vaguely rectangular shape.

[Phredreeke]

Justice for Levelord

[Nfsfan83]

If is it wrong topic please move ( I don't know where I can write it) sorry

I have small sugesstion about Shotgun in NOT POLYMER mode:

I use this settings: https://forums.duke4...tgun-palm-tree/

and this positions for me is perfect:

https://i.imgur.com/Masf60x.jpg

[Phredreeke]

The HRP is dead. Tea Monster and I have been posting things we've done for a successor project. No models are planned for it.

[Nfsfan83]

 Phredreeke, on 25 November 2020 - 10:36 AM, said:

The HRP is dead. Tea Monster and I have been posting things we've done for a successor project. No models are planned for it.

It is model from HRP but better position in game. Can You add this to HRP pack.

[Tea Monster]

It does look good.

[Tea Monster]

 MrFlibble, on 22 November 2020 - 12:15 PM, said:

Looks very good! But since there's manual editing and polishing going on anyway, I think it'd be nice to clean up/straighten those outlines around the photo (I mean the light brown/black angular shape). Not sure why the pixels are so jagged in the low-res original, especially on the left-hand side, but the upscale could have them look less squiggly I think.

If you have the resources, I think many shapes involving straight lines could use some manual touch-up after upscaling and downsampling to 2x as they tend to get sorta blurry. Like the grey metallic thing at the bottom left-ish here. Is the dark blot at the centre of it supposed to be an upwards-pointing arrow BTW? The low-res version looks as if it could be but the upscale just renders it as a vaguely rectangular shape.

This is a big problem. With the Duke Burger sign, the algorithm buggered things up. I wound up literally recreating the shape with a bezier path and then manually repainting the texture into the newly created shapes. The text is recreated with specific effects to match the original tile. The algorithm provided more of a framework for the finished art rather than just touching up what was presented by ESRGAN. In fact, the burger number slot looks off now as nothing was done to it compared to the amount of plastic surgery that was done to the rest of the tile. I didn't touch that as I felt that maybe I'd gone too far for texture fidelity. This is something we are going to have to establish going forward as general rules of how we treat tiles.

The thing is, that once you start down that road, it is the definition of "the slippery slope". Where do you start? When do you stop? Once you have manually edited some tiles to fix glaring errors, will they be visibly un-algorithm-ised to players compared to the rest of the AI upscaled tiles? These are all things we need to decide going forward.

@Nfsfan83 - Yes are posting in the correct thread. Although the HRP is currently dead, there is nothing to stop anyone in the future from resurrecting it in a different form, art-style or under a different renderer. There are still maintenance postings occasionally from various sources. I posted some from work we did on the HHR mod relatively recently.

@Phredreeke I stated that this project is technically a completely different effort with different goals and should have it's own thread so that it isn't confused with the HRP.

[MrFlibble]

 Tea Monster, on 25 November 2020 - 05:58 PM, said:

The thing is, that once you start down that road, it is the definition of "the slippery slope". Where do you start? When do you stop? Once you have manually edited some tiles to fix glaring errors, will they be visibly un-algorithm-ised to players compared to the rest of the AI upscaled tiles? These are all things we need to decide going forward.

I've had enough experiments with various ERGAN models and enough frustration from it to arrive at the following opinion: AI upscaling in its current state (or at least, the kind available via ESRGAN; although the NVIDIA thing that was used for Doom is hardly more advanced) is awesome and has huge potential, but cannot entirely replace a human artist.

Moreover, AI upscaling does not actually guarantee any kind of source material authenticity either, because you can have two or more models that will produce good looking results for the same image, but those results will be at last somewhat different -- while both being, in a way, "true" to the original low-res sources.

But the main problem of course is that ESRGAN does not "understand" what parts of the image it upscales are supposed to be, and introduces all sorts of errors, like warping lines that are supposed to be straight, while adding little to no detail at all that would be added by a human artist.

So my impression is that, just like you said yourself, it is more of a help/framework rather than a tool that will do all the dirty work for the artist. In some cases it turns out to be more of a "stone soup" scenario.

I believe that if you are able to do manual improvements, you should do them. To my mind, if an artist produces a manual recreation of a low-res tile and tries to make it as close to the source material as possible, this is a better cause for authenticity than any AI upscale, especially if the model was trained on something not really related to video game textures or sprites.

[Jimmy]

The reason Doom upscales turn out so good is because 80%+ of Doom's art is very clean. Duke has very inconsistent and gritty art, and this tends to confuse most AI models.

[Tea Monster]

Yep. A lot of photo-source stuff as well, so the AI can get nice and confused as to which algo to use.

[Phredreeke]

I dunno, I think aside from signs and human faces the upscales have turned out pretty good. For the curious, here's a raw upscale without any post-processing done



and the same sprite as it will appear in the next upscale pack - this has not have any manual processing done - it was all automated

[Tea Monster]

That turned out really well. There is some distortion in the fingers, but generally, it looks great. The more I work with these, the more I can see the algorithm at work. I think that this is the best option for getting accurate upscales though. You could go in and fix all the problems, but it would take a shit-ton of work and again, I'm not sure you would be able to correctly fix certain issues without them looking glaring in relation to what you didn't fix, negating the whole effort.

[Phredreeke]

At some point someone more artistically talented than me is needed.

[Tea Monster]

You are doing a great job. The thing is that to recreate everything and expunge all the artefacts, you are probably looking at recreating all the tiles by hand, maybe by using an upscale as a guide for accuracy. I don't think that that is really the point of this project. I think that the odd manual tweak here and there will be great for this. An all-by-hand recreation would be more like "HRP v2" than what we are doing here.

If we keep on as we are going, we will have a great recreation of the original game, but with less jaggies, which is a brilliant thing. Partner that with the voxel pack and you'll have as close to "1996 on steroids" as you can reasonably expect.

[MrFlibble]

 Jimmy, on 28 November 2020 - 12:32 PM, said:

The reason Doom upscales turn out so good is because 80%+ of Doom's art is very clean.

From what I read in the Doomworld topic, there was a ton of manual post-processing involved, including but not limited to manual recreation of all alpha channel masks for every sprite. Ouch. Also the NVIDIA AI introduced a lot of noise that turned out to look bad on the textures in-game (after everyone praised this "gritty" effect when the results were first shown) and had to be cleaned up -- again, manually.

 Tea Monster, on 28 November 2020 - 01:37 PM, said:

Yep. A lot of photo-source stuff as well, so the AI can get nice and confused as to which algo to use.

ESRGAN does not choose the model to be used, it's entirely on the human operator. Also each model consistently produces results based on what it was trained on. By which I mean that every time you process the same source image with the same model, you get the exact same result, no variations. But different models produce different results.

From what I know, ESRGAN works as follows: during the training phase there are two neural networks, one of which tries to create a hi-res image from the low-res source, while the other compares the result to the "ground truth" hi-res original. If results do not pass, the first network adjusts its approach -- I'm guessing based on what the original creators of ESRGAN coded in as possible ways of scaling up images. When enough results end up being okayed by the second model as "close enough" to the ground truth, the model is trained.

 Phredreeke, on 28 November 2020 - 02:57 PM, said:

I dunno, I think aside from signs and human faces the upscales have turned out pretty good.

ESRGAN models do work well, mostly, for pre-rendered character sprites, that is no doubt true. But please take a closer look at your upscale of Levelord for example:

The upper part of the astronaut's helmet is clearly supposed to be rectangular, but on the upscale it has become blurred out and distorted; the lower left part where there is a slight darkening from the proximity of the visor's edge the line has become visibly curved.

The lights on helmet's base have turned into oddly shaped blobs, while I'd suppose they should be at least symmetrical. Surely there'd be not a lot of effort to edit them to be, I don't know, purely circular?

One could of course argue that these are minor issues, and the helmet is the background while the main picture is the face. But these distortions are clearly not consistent with the original source image which was intended to convey something a bit different.

[Mark]

Since showing the upscales in Doom convinced the artists to make changes, should you guys be including an in-game shot along with showing the original and upscale pics?

[Phredreeke]

Sure, it shouldn't be too hard to change that. I did change the jagged border to the "window" on the Bobby upscale after it was pointed out. (as a sidenote, I have a theory of why the edges were jaggy to start with - perhaps the original tile was meant to be larger then at some point they resized it down)

Mind that I was never planning on doing those sprites to start with, just that after doing the faces on the wanted posters I thought I was already half there to do the faces of death ones (and it gave a chance to restore the 1.3D faces)

I'll also point out that I released a new version of the upscale pack. It has much of what was showcased in this topic - although some new inclusions goes against the spirit of this project (previously mentioned liztroop, new textures for Nuclear Winter that were originally unaltered from Duke3D, changing the Atomic text to say Upscale Pack) and most of the world textures were not included (as to avoid having to generate palswaps for those + determine which ones are needed, as I expect the "ERP" to instead rely on indexed hightiles)

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