----------------------------
@Narf:
That one's an archer.
I was just going by what "Daggerfall Imaging 2" had populated in its description field. I guess it makes since that he's an archer and not a fighter mage since he shoots a bow but doesn't have any spell casting animations... lol
----------------------------
@King of Worms:
I think it should be 2x the resolution IMO tho... I will try it. Thank you
Sorry, I first thought you meant 2x the vanilla DF resolution, not twice the AI upscaled resolution (which would be 8x the vanilla DF resolution).
Until then, I use sprites which are cca 1080p in height in case of NPCs.
Here I go exposing my ignorance again, but what does "CCA" mean?
Thinking about it.... I have all the MOBs upscaled with XBR4x times two and then reduced to 50%, than some manual touches to clean up artefacts, plus I corrected skeleton animation phases, used better version of a bear etc.... all with correct full naming... so... maybe you can use that as a base if it helps you?
I think I can use these actually. I'll just pull them from the DREAM mod. See my response to Jukic below.
btw guys, any magic trick to upscale things like this nicely?
The AI messes up text so I clone brushed out the text at the top before upscaling it. I tried several things, but what ended up working best is using MrFlibble's classic xBR softening technique, then running it through SFTGAN. Then I ran it through the unsharp mask filter to accent a few subtle details.
----------------------------
@Jukic:
Jukic wrote: ↑Sat Jan 12, 2019 7:27 pm
@mason: no prob, I wish You guys all the best! I am also "full-in" anyways
Regarding sprites: dark background - no prob. just dont let it be complete black but some specific color, whatever suites You best.
Here are all the orc frames with a funky color background. They're a little rough around the edges. Beware: the AI adds a little bit of noise, even to the background for some reason which may complicate your workflow. It's a shame that the alpha has to be removed in order for the AI to work...
ESRGAN:
https://drive.google.com/file/d/1khYVar ... sp=sharing
SFTGAN:
https://drive.google.com/file/d/1cKnz3o ... sp=sharing
KoW's DREAM mod Orc downscaled to match ESRGAN and SFTGAN 4x scale
https://drive.google.com/file/d/1cKUBbO ... sp=sharing
Below I did a more complex procedure where I composited the results of all three. I think this best combines the strengths of the different upscaling methods (at least in this case). It uses KoW's upscales for the edge of the sprite, then combines some of the details from both ESRGAN and SFTGAN.
Basically, I used KoW's orc from the DREAM mod (scaled back down from 8 times to 4x the vanilla DF resolution) as the bottom layer and used it for the edges. I selected its outline, grew the selection by 2 pixels, feathered by 3, then deleted that selected portion from the ESRGAN and SFTGAN layers. Basically, I wanted the outline of KoW's orc with the inside texture of the AI upscaled stuff. I then blended the ESRGAN and SFTGAN by setting the ESRGAN to the top layer and setting its opacity to 50%. SFTGAN was the middle layer with 100% opacity. I have no idea how feasible all that is to do with batching. Feel free to experiment with your own methods.
----------------------------
@MrFlibble:
Training models is beyond my capabilities ATM, 'cause I'm stuck with CPU mode even though I have an NVIDIA card but my current distro is officially unsupported by CUDA.
That sucks!
I think training is also outside my capabilities. From what I understand, it would take my GTX 1060 over a week to train a model even with CUDA... although I could be wrong about that.
On another note, I discovered that if you use Sinc interpolation instead of pixelise + nearest neighbour when applying xBR softening, you get sharper results with all the advantages of smoother lines.
So when you use the Sinc interpolation, you don't apply the pixelize filter at all? I tried that but I'm not quite getting the same results. One caveat is that I'm now running Gimp 2.10 which replaced the Sinc interpolation with two other algorithms, NoHalo and LoHalo, which are
supposed to be basically the same but "better". Not sure how.