Dune II: The Neural Upscale

[MrFlibble]

So the increasing the resolution of old video game graphics using neural networks is all the rage these days, so I decided to jump in!

My first try was waifu2x, which turns out to be not the best thing around by today's standards, yet it produced results far beyond what a conventional pixel scaler can do. But now there's ESRGAN, and people are getting pretty interesting results with a model trained on Manga images.

So I tried this out, with quite interesting results (IMO), and I couldn't resist doing some Dune II screenshots too (all coming from the official Westwood screenshot pack):

None of the images were edited in any way, except scaled down from 1280x800 to 640x480 (for correct aspect ratio) and converted to the original 8-bit palettes using mtPaint with Dithered (effect) dithering method.

[X3M]

The missile tank looks a bit weirder now. But overall, it looks more smoothly.

Have you some more images. Like how the tanks look like in mentat? Or like how they look like in the heavy factory selection menu?

Edited January 12, 2019 by X3M

[MrFlibble]

Hey, sorry for not replying sooner.

I was experimenting with different approaches to this, including neural network interpolation and using another network called SFTGAN. It does some pretty good job with screenshots from Command & Conquer (as shown here, scroll down for SFTGAN results).

[X3M]

No problem, keep me updated. It looks like fun.

That medium tank in the first picture, has a weird looking cannon.
But those (darker) rocks, now they really look like rocks. I though they where bushes or something in the oldest versions. Unless,... they are bushes.
It looks good! But needs some cleaning, here and there I think.
 

Seems yet another forum that I never knew of.
But seeing that Plokite Wolf is active there. I will only lurk.

[MrFlibble]

In the meantime, more advanced models have been trained for ESRGAN, here's a little sample:

The only drawback is that is fades colours quite a bit, I had to brighten them up in GIMP.

[X3M]

Same pictures, but subtle changes.
Nice work!

 

I am curious if it is all done. To get my hands on this version.

This is how a remaster should look like

[Nyerguds]

On 3/17/2020 at 2:22 PM, MrFlibble said:

The only drawback is that is fades colours quite a bit, I had to brighten them up in GIMP.

That, and neural networks trained for images still have serious issues with text, lol.

[MrFlibble]

22 hours ago, Nyerguds said:

That, and neural networks trained for images still have serious issues with text, lol.

Well, you can't have everything at once Besides, usually you don't need the upscaled fonts anyway, not with the neural networks at least.

On 3/22/2020 at 3:19 PM, X3M said:

I am curious if it is all done. To get my hands on this version.

Not sure if I got what you're saying here but if you mean to ask if I processed the entire set of Dune II graphics, no I did not, I just upscaled several screenshots from Westwood's preview. But it is certainly possible to do that. You can try out ESRGAN yourself, even if you don't have an NVIDIA card it can run on CPU, albeit much much slower. My NVIDIA card is too old for CUDA anyway so I had to use the CPU as well (duh!). There's a guide on how to set it up here.

I think I forgot to mention what models I used to get the latest results. The two are rebout_interp and Spongebob v6 De-Quantize, both available from the Upscale Wiki. I interpolated these two models (with the net_interp.py that comes with ESRGAN) at alpha = 0.8 (meaning it's 80% SpongeBob and 20% Rebout).

Note that these models are to be used with the "old architecture" ESRGAN, available here.

Here are some more results I shared elsewhere (not from Dune II), with the same SpongeBob/Rebout interpolation. I really like how it processes images:
https://imgsli.com/MTMzMjE

https://imgsli.com/MTMzMjI

https://imgsli.com/MTMzMjM

https://imgsli.com/MTMzMjQ

https://imgsli.com/MTMzMjU