AI Neural Net upscaling of old pixel backgrounds

Hopefully I’ll get to MI2SE later… =)

The first iteration is complete. All upscaled backgrounds were converted and saved as “layer0”. Unfortunately some SE rooms consist of more layers and other additional objects so they (would) require hand tuning.

If you want to test it just unzip the archive to your MI1SE folder.
https://drive.google.com/file/d/1bPy7XtihcKr7tsiunx2_J_vvF75OnVvc/view?usp=sharing

1 Like

This is something! Pretty cool! Now we can just repaint those parts to better quality :slight_smile:

Have you checked, just in one room, if all objects/actor hotspots are still (more or less) in the right places? The Scumm bar seems to be a perfect testroom for that.

I really wonder who exactly do you mean by “we”… :wink:

Anyway, any person who will try to do it will need both original VGA images and MI1SE images scaled to the same resolution (i.e. the height of 1056 pixel). I already had the upscaled and resized VGA images but the MI1SE images were still split to chunks (1024x1024 pixels) so today I had to create some other scripts and merge the chunks together.

Both packed folders can be downloaded from my Onedrive:
https://1drv.ms/f/s!AqY4NY4gpRQkg0ZQjHqM8BSDTjM2

Unfortunately some SE backgrounds are a bit expanded or something - e.g. VGA version of “fork” is 2880 pixels wide and MI1SE version is 3072 pixels wide…

Edit: Oh, now I can finally see the black border on the right side so maybe it’s not a big deal. =)

1 Like

It’s exactly “more or less”. :smirk:
Many objects in the bar are slightly moved from their original positions. The differences are not big (probably up to 20 pixels) but they are noticeable.

luckily @patrik3dspacek will do all the cleanup by redrawing the backgrounds :wink:

3 Likes

I’m really glad Patrik will take care of it. :wink:
But I guess it would be more practical to split the work among several people because the number of backgrounds is really big.

After a few conversion cycles I became tired of using Monkey Island Image Converter because it requires mouse and every time I had to set the target path so I downloaded its source code (unfortunately it’s in Delphi) and turned the MI1 part to a C# console application that could be easily run from DOS batch files. It really helped me to speed up the conversion process.

The “dds-dxt-mi1.exe” and my conversion batch file can be downloaded here:
https://1drv.ms/u/s!AqY4NY4gpRQkg0dFzjwCoOGAcI5z

1 Like

My list of all background images with comments - I’m still verifying sizes and transparency…

1 Like

There is a question; what version would be better to repaint? classic or SE? if classic BGs doesnt fit with animations sprites of SE… I guess, it needs to be painted based on classic art with those new SE sprites coordinates.

Will there be a problem to do this with MI2SE as well?

I haven’t tried to replace any additional objects yet but I plan to place all the available graphics elements of both versions into one multilayer image (one file per room - I think ImageMagick could handle it too) and try to verify the locations of objects. The problem is that some SE objects does not have visible border so it’s hard to predict if the old sprite will fit or not.

I suppose the best way would be starting with a combined versions of VGA and SE background…
(E.g. I think there should be some ships in docks but there are none in VGA version)

Anyway, we need to start with a few testing rooms and see what works the best.

Edit: Damn, conversion to PSD seems to work but the PNGs exported from MISE Explorer doesn’t include alpha channel. I’ll have to modify my scripts and run them again - this time on original DDS files. Hopefully I’ll get to it tonight.

1 Like

I use PS for all…you can install Nvidia tool, that will add DDS support to PS. DXT5 w alpha. Works well.
I guess, you can open BG in PS, place the sprites as you can see in the game and the coordinates are in “info” panel. Then you can calculate pixels difference.

PS does support DDS but my image viewer does not so it was more convenient to use PNGs for exploring the folders.

So you are stitching all the background chunks together manually? I’d like to avoid as much manual editing as it’s possible because it gets boring. If you are interested here’s my latest “repack” - all the room graphics merged to PSDs (layer 0 = upscaled VGA, layer 1 = layer0 from SE, … extra layers and objects from SE)

https://1drv.ms/u/s!AqY4NY4gpRQkg0kgL2FzH60coAX4 (fixed link)

I guess, you can open BG in PS, place the sprites as you can see in the game and the coordinates are in “info” panel. Then you can calculate pixels difference.

Sure - it’s rather easy to calculate the transition vectors but many SE objects don’t have visible (solid) border so it’s not clear how big is the rectangle that’s being copied to the final game scene. Even if you place the revised object on the right spot the margins could be different.

I did some experiments with Lookout background but I just could not find a way how to improve the upscaled VGA background and keep its style at the same time. I found myself editing the SE version version instead and I made it a bit … um, darker. =)

(By the way, the gate in VGA version is heading to an illogical direction and both versions share badly build arcs - they would immediately collapse in real world.)

Upscaled VGA (two layers)

SE

Modified SE

2 Likes

twp1_cartoonsmooth
twp2_cartoonsmooth
Thimbleweed Park upscaled with AI easy to do
(full size 1712x960 up-scaled)

I think that proves the original was perfect as it was. The AI has a weird way of changing the text into something more “plausible”: bus pass? No, that surely must have read “bug pogs”!
Redeem bottles herb. Mail Brad, because he played too many conto.

3 Likes