I very much agree with this post.
Forget party hats, Christmas tree hats all the way man.
Nice! I should really dig up the lua scripts I wrote and post them. But I probably won't! Cause that's just the kind of guy I am this week.
Output
I'd make an AviSynth script for this if I understood at all how its indexing system worked.
Of course, it's notable that my method is definitely different to Randil's, and certainly worse.
turned out like this:
I think the problem lies in the distance function. I'm using the same one I used for the mosaic script I made, namely this: If the pixel you are to choose a color for has RGB values R1, G1 and B1, and color i (i is 1,...,15 since Mario Paint has 15 colors) has RGB values R(i), G(i), B(i), then the distance function is defined as
d=|R1-R(i)|^p + |G1-G(i)|^p + |B1-B(i)|^p
Here the RBG values are between 0 and 255. The script loops through all 15 colors and picks the one that gives the smallest d.
p is a parameter you can change in the script. I used p=1 in the pictures I posted.
I seem to remember Derakon mentioning in the Mosaic thread that you should weight the RGB values differently. I might need to do this. Another option would be to choose an entirely different distance function. Any thoughts on this matter?Neat! Now add dithering. :)
We have the same greyscale issue, it appears.
I've figured out a way to convert video files using only the palette, but I would really rather use the algorithm that Randil has got going on here. A configurable AviSynth plugin would be unlikely, but wonderful. :p
If anyone wants, I can produce a GIF/YouTube video, but I can't promise it'd be too pretty.[...] but I would really rather use the algorithm that Randil has got going on here.
Good luck!I think the problem lies in the distance function. I'm using the same one I used for the mosaic script I made, namely this: If the pixel you are to choose a color for has RGB values R1, G1 and B1, and color i (i is 1,...,15 since Mario Paint has 15 colors) has RGB values R(i), G(i), B(i), then the distance function is defined as d=|R1-R(i)|^p + |G1-G(i)|^p + |B1-B(i)|^p Here the RBG values are between 0 and 255. The script loops through all 15 colors and picks the one that gives the smallest d. p is a parameter you can change in the script. I used p=1 in the pictures I posted. I seem to remember Derakon mentioning in the Mosaic thread that you should weight the RGB values differently. I might need to do this. Another option would be to choose an entirely different distance function. Any thoughts on this matter?
I double checked that the program works correctly, and it seems like it does. Strange...
Old version:
Lab version:
I'm not sure which one is best of those two.I very much agree with this post.
Forget party hats, Christmas tree hats all the way man.
Not needing to dither would be better, but there are so many colors missing from the palette that you probably need to either use extensive dithering or be extremely picky about which source image to use.
As for actually quickly creating images like this in-game... I would guess that the fastest way is to use the "draw sprite" tool to draw 16x16 blocks of solid color (or the eraser for the almost-white background color), then switch to smaller tools like the 2x2 pen and (if necessary) custom stamps or carefully-planned spray-paint for the detail work. Actually, custom stamps aren't that slow to create so they might help a lot. I would avoid the paint fill tool since it seems to be very slow, although maybe it's still the best way in certain cases. And the fastest way to fill the entire image with a solid color seems to be the "new page" button, but unless the vast majority of the image is a single color, it's probably faster to go straight to 16x16 blocks at the start.
Also, here's an idea... you could greatly increase the number of colors that appear to be in the palette for any 120x80 block of pixels in the image, by using the animation tool at speed 13 to create a 60hz flicker effect to simulate having colors halfway in-between any two colors in the palette. Compare the region with Mario's face in this image with the first one above:
EDIT: Some more images with this approach:
As you can see, it still has some trouble with smooth transitions, but I think it looks better now.
EDIT2: One more to show the differences - First one is the one made with my first script, then comes the 2x2 version, and lastly the original:
Flygon's picture made with the 2x2 script:
I made another approach where I divided the picture into 2x2 blocks and looked for the best combinations of 2 colors (one color in the NW and SE corner, another in the other two corners) for each of these blocks. This took care of the grey-scale issue, but the resolution isn't very good now
I don't know how much time it takes to apply the scolorq filter
It varies, but scolorq took 11 seconds on the color scale image on my system. I think scolorq can be optimized more as well (it seems to spend about 20% of its time on allocating and deallocating memory on the heap in inner loops, and about another 60% of its time on simply inserting elements into STL vectors, which is a lot of time to be spending on the bookkeeping instead of the actual computation).
In sequence: Naruto; a car; Megaman; a manga drawing; Mario paint cover crisp and neat; Mario paint cover with more accurate colors.
Once it finishes an image, it immediately starts painting the next one over the previous one.
Here's my latest WIP. I don't remember what it exactly contains. IIRC the script got stuck somewhere.
http://dev.w3.fi/%7Ebisqwit/mariopaint-wip-unknown.smv
Oh, and about dithering: All these images are converted with Animmerger using Yliluoma-2 dithering at matrix size 16x16 with maximum 4 color combinations per pixel, maximum precombined colors, assumed gamma 2.0. The original was resized with lanczos filtering and sharpened minimally prior conversion. The dithering was rigged to produce patterns aligned to Mario Paint's brushes where possible.
I've been secretly working on a Mario Paint drawing TAS this year.
Why use ordered dithering when scolorq or even Floyd–Steinberg would create a visually much better result?
