Images generated by RastaConverter

[AtariNerd]

I might hazard a guess, that if you were to set the number of solutions incredibly high, maybe even up to one solution for every pixel, you might theoretically eliminate any artificial "wall" and it would slowly improve for however long you let it run. Over the process of billion of evaluations, it could move further along (or at least have better dithering, with less banding, if those options are used), than with just a handful of solutions, but the grinding time, especially at the beginning, might try most peoples patience,

 

My PC is much too slow for such attempts, though and patience perhaps not sufficient enough, but I might try it someday, anyway, if there were an image I really, really wanted a certain result with, like a portrait of someone/something I cared deeply about.

[+MrFish]

Probably when it is still running, it's a 64-bit variable. Somewhere during saving or loading it gets assigned to a 32-bit variable and causes and integer overflow.

 

Sounds plausible.

 

I'll try and just leave my secondary machine running.

[+MrFish]

I might hazard a guess, that if you were to set the number of solutions incredibly high, maybe even up to one solution for every pixel, you might theoretically eliminate any artificial "wall" and it would slowly improve for however long you let it run. Over the process of billion of evaluations, it could move further along (or at least have better dithering, with less banding, if those options are used), than with just a handful of solutions, but the grinding time, especially at the beginning, might try most peoples patience,

 

My PC is much too slow for such attempts, though and patience perhaps not sufficient enough, but I might try it someday, anyway, if there were an image I really, really wanted a certain result with, like a portrait of someone/something I cared deeply about.

 

Well, I think the first posts by Gunstar showed it was worth exploring further, and now, after letting them go this far and seeing that progress continues at a good pace, I think it's worth taking to the next step. I'd planned to use 10,000 solutions and see what happens. For me, it doesn't matter how long it takes because this is a secondary machine. Even on my main machine, if I set it to asingle thread, it barely interrupts anything. I just check on them once or twice a day to see how things are going.

 

I was thinking to stop the current ones at about 30 billion, but we'll see how things work out. It's really made some big improvements over time in most areas. I'm amazed at how well the details are matching up with the destination image.

 

On these two, I'm really looking to see more improvement on the lower part of the left hedge where the yellow in the field starts to kick in and corrspond with it horizontally. With these two versions it seems to be the main areas where things haven't progressed much. Everywhere else, things are really getting dialed in.

Edited November 4, 2017 by MrFish

[Sheddy]

I can't understand the code, but from previous explanations, i believe the random start points of many solutions helps it avoid finding the easy to guess(highly probable to keep hitting randomly) solutions that don't actually work out better long term. That's important as it can't really undo earlier choices because they will not be better.

Edited November 4, 2017 by Sheddy

[Gunstar]

From my experience, I did start getting better results the more solutions I used, but being clueless about the inner workings of Rastaconverter and the algorithm, I was afraid to set it too high. early on I left it in default (1?) and then went to 10. Later I started hovering around 40-50, and finally settled on 42, being a Hitch-Hiker's fan. In the last month I decided to raise it to 128. simply with the logic of 128 color palette and the possibility with solutions having something to do with the colors chosen for the output image. All just guessing on my part of course since I'm clueless to the program's "thought process" of it's algorithm.

 

But early on I was also hesitant to adjust anything on the GUI too much, until I realized that some of it, like contrast, can be adjusted quite drastically and not ruin the image, but make it easier, seemingly, for choosing color and luminance levels when down scaling from 16 or 24-bit color to 256 and then a 128 color palette, which the program on average only chooses about half the palette for the output image, So basically a downgrade to 64 colors plus all the other Atari graphics restrictions with cycles for sprites and use of DLI's.

 

But in all of this, I still stand by my belief that it's possible to get fantastic results in less than 100 million evaluations, and this is useful to avoid the line errors that show up when viewed on real hardware, ruining the image. None of the pictures I've posted after maybe the first couple dozen have gone much over 100 million evaluations, most actually are still around 50 million evaluations. Can they be improved with billions and other pre-editing like Mr.Fish is doing? Sure, if you want to spend that much time on one image, but the ratio of improvement, is such a small percentage I find it of little use. Many of the more recent pictures, the last 10 or so, have only run for 10-30 million evaluations, because I am getting better and better at setting the color levels and I also believe due to going to 128 solutions. So I still say that equally impressive results can be achieved without billions and billions of evaluations, and it's all in the GUI settings, not how long it runs.

 

I will start really upping the solution number far beyond 128 and see if it helps me more.

Edited November 5, 2017 by Gunstar

[Gunstar]

I inverted a couple words above, I meant the percentage of improvement is such a small ratio to amount of evaluations. (beyond 100 million)

Edited November 5, 2017 by Gunstar

[Gunstar]

The first of some automobile pop-art.

 

Cruiser. 88 colors.

 

General Lee. 57 colors.

Cruiser.xex

GeneralLee.xex

[Gunstar]

Classic lines. 10 colors.

 

Porsche Carrera. 8 colors.

 

Sweetness. 18 colors.

Classiclines.xex

PorscheCarrera.xex

Sweetness.xex

[Level42]

Here are some pictures of Amsterdam.

 

24M evals

output.png-src.pngoutput.png

 

13M evals

output.png-src.pngoutput.png

 

20M evals

output.png-src.pngoutput.png

 

All with laoo palette and chess dithering and /s 1

 

I have a pretty slow computer, so perhaps somebody could run a couple of hundred million evaluations with better settings

 

Blegh....Amsterdam. Always Amsterdam.....tourist infected and ruined....

 

Use some nice pictures of a really great city instead:

 

https://www.derooijfotografie.nl/11-mooiste-fotos-rotterdam/

[Gunstar]

Chevy V8. 62 colors.

 

Hot Rod. 65 colors.

ChevyV8.xex

HotRod.xex

[Gunstar]

The first of some Steam-Punk/Diesel-Punk art.

 

Diesel-Punk Locomotive. 51 colors.

DieselPunkLocomotive.xex

[ivop]

Blegh....Amsterdam. Always Amsterdam.....tourist infected and ruined....

 

Use some nice pictures of a really great city instead:

 

https://www.derooijfotografie.nl/11-mooiste-fotos-rotterdam/

Wow, over five years old But next time I do some Dutch images, I'll be sure to include all suppressed minor cities

Edited November 13, 2017 by ivop

[a8isa1]

Real Colors!

 

 

Here's a screenshot of a conversion session, very early in the process.

 

OK. Back to reality.

 

I did generate the above picture conversions using rastaconverter but I used palettes with vivid colors. The pictures are displayed using the same palettes.

 

Here is what some of the pictures look like using the "default" palette for Atari800

.

Atari800

 

The colors are muted but still completely present.

 

I stumbled on this technique which yields more saturated colors from rastaconverter by accident. I was reading Trebor's thread regarding 7800 palettes and learned that palettes were created to make emulators present video like the real machines they emulate. They are biased. This made me wonder if we should be using the same biased palettes when converting pictures with rastaconverter. Shouldn't we use a neutral palette?

 

I don't have Trebor's expertise and don't have any tools of my own to speak of so I used Altirra and Jakub palette (which looks fairly balanced to my untrained eyes). I adjusted the start hue going on Analog Pocket Programmer's Guide's description of the hues. Hue 1 Light Orange. Hue 2 Orange. I arrived at Hue Start of -44 degrees. From there I assumed colors were separated by 22.5 (360 / 16)degrees. Yes, I wondered (and still do wonder if grayscale is a special case that should be discounted but I ignored the idea). Altirra's edit colors menu has a granularity that forced me to select 22.4 degrees. I made brightness 0. Contrast is now 80%. I don't remember my rationale for selecting this. Pictures seemed to be more agreeable when running rastaconverter so I stuck with that number. I was getting some nice color but not any completed conversions. At some other point I decided I wanted the color red to be more vibrant. I knew this is a weakness in Atari's rendering of color but I tried to make it just a little better. I boosted Altirra's saturation number to 30. I exported these changes to a palette file.

 

Running conversions with the new palette I noticed one thing instantly, the source and destination pictures in Rastaconverter's progress window now had similar amount of brightness, contrast, and absolutely saturation. Not long after that I also noticed a recognizable picture was forming within seconds (/s=1 parameter was used). I also saw more colors and the dither pattern was not prominent. I some pictures I can't even find the dithering.

 

I get pictures that are strong facsimilies of the sources in just minutes. (again '/s=1').

 

Now for the bad news. The pictures stall very quickly. In my mind when normal dist is ticking off slowly in the 5th significant digit the picture is "stalled". They stall with quite noticible errors still in them.

 

I can't produce what I consider completed conversions using this technique. I've run some conversions for up to a day. Most of them don't even reach the quality of the ones above, using /s=1, which were all done in less than 5 minutes. In fact I set the conversions with /max_evals=41000000, 41 million.

 

Anyway, I just wanted to share what I learned. Below I'm enclosing the two palettes with vivid colors and two ATRs containing my executables of the pictures above. For the PAL executables I cheated. All my original conversions were done using the NTSC palette. For PAL conversions just grabbed the OUTPUT.PNG file, which is already a local solution, and fed that back rastaconverter but using the PAL palette. In this run I skipped dithers, and image settings. I assumed rastaconverter would resolve the pictures instantly. They didn't. However, after just 4.5 million evaluations they compare favorably with the NTSC ones so I'm running with the decision

[EDIT²]

 

How I use these palettes in atari800 (linux build)

 

./atari800 -ntsc -paletten <palette path> -windowed -nosound a8isa1_set3.atr

 

For Atari800Win Plus I think equivalent parameters are available when the program is evoked from the Windows Command line.

 

I don't see an option to import palettes in Altirra but the video color adjusting tool will let you make the same changes I described above.

 

For PAL Everything is the same except Hue Start is -15

 

For either video standard start with the jakub.pal preset

 

[EDIT]

These pictures work for me on my 800 XL but my CRT (VGA on a scan converter is failing). I'm curious how (or if) they work on real Ataris for people with better equipment, preferably via a PVM. I don't want people to mess up their preferences but if these work they would need monitors at their factory defaults (Atari likewise though I don't know how one restore an Atari to the factory setting for the color trim pot)

 

I don't think these will work well via VBXE which I understand has the laoo palette installed but all feedback is welcome.

 

Thanks.

 

-SteveS

 

Palettes & ATRs

a8isa1.zip

Edited November 15, 2017 by a8isa1

[Gunstar]

For a split second I thought you were showing screen shots of an Atari with a video upgrade, because that's what Rasta images look like to me on my 1200XL, but then my eyes adjusted to the blinding color, and I noticed the Atari 800 Emulator bit. Very interesting stuff, I'll definitely try your palettes and check out those images .XEX's on my 8-bit and see if I'm blinded by science.

[a8isa1]

For a split second I thought you were showing screen shots of an Atari with a video upgrade, because that's what Rasta images look like to me on my 1200XL, but then my eyes adjusted to the blinding color, and I noticed the Atari 800 Emulator bit. Very interesting stuff, I'll definitely try your palettes and check out those images .XEX's on my 8-bit and see if I'm blinded by science.

I have and old TV I can move back to the desk. It's default setting already has colors that are too saturated. My eyes might fall to the floor.

[Gunstar]

Steam-Punk Girl. 28 colors.

 

Steam-Punk Dirigible. 30 colors.

SteamPunkGirl.xex

SteamPunkDirigible.xex

[Heaven/TQA]

Still amazed by the pics esp steam punk.

[Wrathchild]

 

I can't produce what I consider completed conversions using this technique. I've run some conversions for up to a day. Most of them don't even reach the quality of the ones above, using /s=1, which were all done in less than 5 minutes. In fact I set the conversions with /max_evals=41000000, 41 million.

 

 

What was the effect of trying different /predistance and /distance models?

[a8isa1]

 

What was the effect of trying different /predistance and /distance models?

I generally only use YUV. CIEDE when the colors don't seem right. Euclid when the colors of the source aren't natural, computer generated images, art (some), that sort of thing.

 

I didn't know distance/predistance could affect the progress of renderings.

[Wrathchild]

if the source image only contains pixels of rgb value matching those in your palette then any color distance method should find those colours as the distance is zero.

If not, then I would expect CIEDE to be the better at identifying the 'nearest' colour in the palette. CIE94 can be used to process faster.

[RAM]

a8isa1, That's how I remember my Atari looking on a colour TV, very vivid bright colours, not the washed out colours that emulator palettes show (especially red). If you wanted desaturated colours, you turned the colour knob leftwards!

[a8isa1]

a8isa1, That's how I remember my Atari looking on a colour TV, very vivid bright colours, not the washed out colours that emulator palettes show (especially red). If you wanted desaturated colours, you turned the colour knob leftwards!

 

I dragged out the old "portable TV". Sure enough it is as colorful (but not the color red) as the screens I posted.

Edited November 16, 2017 by a8isa1

[Gunstar]

 

I dragged out the old "portable TV". Sure enough it is as colorful (but not the color red) as the screens I posted.

My reds are as good, but then I have a special video upgrade. Like my 'Chevy V8' pic above, compared to viewing it on my real Atari, that .png image does look drab and almost brownish. My Steam-Punk pictures with all their subdued colors are perfect for stock Atari's.

Edited November 16, 2017 by Gunstar

[a8isa1]

My reds are as good, but then I have a special video upgrade. Like my 'Chevy V8' pic above, compared to viewing it on my real Atari, that .png image does look drab and almost brownish. My Steam-Punk pictures with all their subdued colors are perfect for stock Atari's.

Unfortunately I don't have any PAL gear. I'll look at your pics again with one of my modified palettes.

[RAM]

Unfortunately I don't have any PAL gear. I'll look at your pics again with one of my modified palettes.

 

I had a PAL TV and Atari, so that might explain it. I remember reds being very red!