Daggerfall Workshop Forums

Hi! Just briefly:

I can try to pinpoint the most offending ones

269 and 264 mobs are different. One has tusks and the other one does not. And there are other minor differences between them.

About the outlines - Yup, I mean feathering. Yes, its confusing, technically it should end up as a hard outline... but Ive noticed my outlines are softer than the ones you produce. But it might be because of other reasons than feathering. I must check it more in depth.

I will create some comparisons between 4x and 8x ingame so we can decide it this has any value...

Ok, here's my "fast" 4x-8x conversion https://files.fm/f/bj6hnzxy

it's for sprites between 256-511, I'll process the rest later

Ok, here's my "fast" 4x-8x conversion https://files.fm/f/bj6hnzxy

Thanks Phredreeke! They look good so far.

269 and 264 mobs are different. One has tusks and the other one does not. And there are other minor differences between them.

Ah, I see it now. They also have different feet and tails. I've got the Wereboar (269) upscaled now.

Ive noticed that the Seducer mob is missing. Thats mob 284...

I've now got the seducer upscaled and its forward facing frames sharpened a little.

My goal for this week is to get the sharpened forward facing frames "fixed," upscale the ghosts, and use Phredreeke's masks to refine the outlines on all the other frames, then I'll upload the whole set again as one unified pack.

Wow, thats awesome, Im looking forward to it.
Im currently trying to somehow merge the old textures/materials with the updated version I received from Quasifex. Its basically PITA, I expected something different but Im still glad for everything he did for the case. Plus he told me I messed up my normals, because Unity uses OpenGL format and I created everything in DirectX format, so... yeah.

Better luck to you guys

...and here are the rest of the "fast" masks http://s000.tinyupload.com/index.php?fi ... 9823315487

...and here are the rest of the "fast" masks http://s000.tinyupload.com/index.php?fi ... 9823315487

Awesome, thanks again! These "fast" masks are leaving a feathered edge that have to be cleaned up, but it's a super easy fix with a batch-able script so it's no problem at all. The end results are looking good!

With any luck I should have all the MOBs cleaned up and will be ready to start working on the ghosts by tomorrow night.

Plus he told me I messed up my normals, because Unity uses OpenGL format and I created everything in DirectX format, so... yeah.

I think it's a pretty easy conversion between the two; just need to invert one of the color channels, but I don't remember which. Hopefully it won't be too big of a deal!

Great news!
And yes, Ive found out I can invert the green chanel and that will basically convert the normals from DX to OGL format.
I guess I can create a automation for it. So thats a good thing.. ufff
Now I will have to go thru the 1000 normals and decide which are in DX and which in OGL format... because its a mixed work of me (DX) and Quasifex (OGL)... nothing is easy here. But Im used to.
Good luck!

Hi guys.

I've tested the upscaled textures and mobs and must say I was blown away. This is such a huge deal and when I fight some mobs I barely see the odd occurrences in the upscaled results. I also like that the textures have a slight noise effect, which makes it gritty and more 'realistic' instead of too cartoony

Great work!

I've tested the upscaled textures and mobs and must say I was blown away. This is such a huge deal and when I fight some mobs I barely see the odd occurrences in the upscaled results.

Thanks! It's always great to hear that people are enjoying our work.

In the meantime I expanded upon the xBR/xBRZ multipass method I described earlier. As the up- and downscaling iterations resulted in blurring, I tried to counter that with sharpening the source image prior to scaling it with xBR(Z). This produced pretty decent results. I also came to the conclusion that my earlier idea to blend the ESRGAN upscale with an xBRZ upscale wasn't that bad after all.

The method works with both hand-drawn and pre-rendered sprites and seems to produce better results than anything that I tried previously:


It also does a decent but not outstanding job with character portraits:

I used a different interpolated model this time. A funny thing though, I mixed up the net_interp.py script and its backup copy that I had created, so the model I ended up using is an interpolation at alpha = 0.8 of two previous interpolations: ESRGAN/Manga at 0.2 and ESRGAN/Manga at 0.5. I discovered this just now. Duh.

Anyway, the pre-processing multipass method goes as follows:

1) open the original image in GIMP and scale it up to 400% using nearest neighbour; scale back down to origginal size with Sinc, and export the result

2) open the result in Scaler Test and scale up 4x with xBR; save

3) open the xBR upscale in GIMP, apply Gaussian blur at 1 pixel radius and scale down to original size with Sinc; export

4) scale the result of the above steps with xBRZ to 4x and save

5) repeat step 3 with the two-pass xBR/xBRZ image you just got

6) repeat steps 4-5

UPD: The actual Manga/PSNR interpolation gets more blurry results that I don't like, but the simple ESRGAN/Manga interpolation at 0.5 produces essentially the same results as the above.