Pixel art and classical art (palettes again)

[Tourist]

I'm still wrestling with color choices and could use some help.

Here's a piece cropped from Calinerie by William Bouguereau. 



You can see the whole thing here (about halfway down the page:
http://www.artrenewal.org/asp/database/art.asp?aid=7&page=3

I admire the look of classical oil paintings like this.  I can't seem to translate this into a limited palette suitable for pixel art.  I'm not trying to recreate this piece, but I do want to figure out a palette with the sort of glow in the faces above. 

Too few colors (8 or so) and there is either too much contrast or too narrow a range of hues.  Too many (12+) and I either get banding or have to enlarge my doodles.   

Any ideas? 
Tourist

Btw, the ARC site (the one linked above) is a fantastic resource for browsing artwork.

[Gil]

Notice the hue shifts. Colors don't ramp from bright to dark only, but switch from reds to more greenish hues too. As far as I can tell the hue is correlated to fat tissue here. Forehead, temples and jaw get more green because there is less fat tissue between the bone and skin. The parts like the cheeks, nose and chin get more reds, which gives it a highlight that seems to be radiant from within. It suggest some sort of subsurface scattering, which makes the skin tone so vibrant. Notice that the darker parts with a lot of fat have again some brownish red inside.

[blumunkee]

There is a deceptively large amount of yellow in the image. It is the main highlighting force. Even so, the reds are strategically placed and draw the most attention.

There are a lot of grays and green grays in the painting. The neutral shades are used as buffers between the brown shadows and the yellow highlights. This is especially visible on the side of the child's face. The colors in the flesh and cloth are actually quite muted, but the grays help pull them out to give a glowing effect.

Understand that the optimal palette depends entirely on the size of the piece. If the size is greatly reduced, say to 100x100 px, 16 colors would probably be optimal. If the piece where at the same size as the snippet here, 32 or even 64 colors might be optimal.

The light scattering effects suggest dithering would come in handy. The palette would need to be designed with this in mind. In order to dither properly, you need colors that are close enough in value that they blend well. If the colors are too contrasting, the dithering will look harsh and distracting. If they are close enough in brightness, they'll create a nice blending and texturing effect.

It's worth emphasizing that dithering does create texture. Many beginners use dithering as a poor man's gradient tool, typically with poor results. In a piece like this, the texturing side effect would probably be more useful than the blending main effect.

[Tourist]

Thanks, terrific analyses.  That soft, sub-surface scattering look is what I'm after.  And thanks for the suggestions about the number of colors, that gives me a rough target to shoot for.

I have to say though that there aren't any gray-greens in the image.  It's a yellow brown with very low saturation.

Trying to build a palette still eludes me.  Here's just the forehead of the figure on the right.



Here's an approximate forehead, 32 pix across.  It doesn't look smooth or glowing or soft or anything really.  5 colors of skin tone.  I chose 32 pixels to make these quickly, it's not a firm restriction or anything.



Five colors might not be enough.  I add a blend between each of the 5 color blobs, making 9 colors total, not counting the hair.   Whatever number of colors in the forehead, the same number can be expected for the pink cheeks.  Plus 3-4 for the hair and another 2 for AA on the hair.  And red for the lips, white for the eyes..... successful 9 colors across the forehead translates to about 32 colors minimum just for the face, I think.



Not too bad.  The left side of the forehead blends nicely.  The right side not so much.  Maybe dithering on the right (the poor man's gradient tool)?



That looks terrible!  My dithering skills are pretty weak though. 


I think this shows that blending works for colors that are very close to each other, but it's not good enough for colors that are moderately different.  I couldn't get the dithering to work well at all.  For comparison, here's what the original looks like when resized to 32 pixels across.



It loses a lot, but it still looks so much better than the 9 color job. 

So what do you all think?  Do any of these approach the soft shades that the original full size image does?

Thanks,
Tourist

[Tourist]

Doubled in size to 64 pixels, went to 11 colors and then dithered like mad.



The lighted side looks fine, the shadowed side isn't quite there yet.  Are there different guidelines for dithering shadows vs highlights, or are the colors just too far apart?

Does the dithering work?

Edit: more details on the technique.

Here is one artist creating an oil painting in a classical style.  He starts with a sketch and then adds highlights and shadows.



What I thought was interesting is that he uses a dark red for the shadows and a blue that is almost white for the highlight.  My color picker shows these to be 180 degrees opposite in hue, so the artist gets a strong contrast without using monochrome grays.

Then he layers on paint on top of this for the skin color and so on.

In trying to replicate this, I used the dark red to blue-white and reverse engineered the light, midtone, and dark colors of the piece in the original post.  It looks like roughly 45% of the red/blue layer plus 55% of the skin tone, and then additional red on top.  The red and green components were almost in the exact 45-55 proportion, and the red as a 70-30 mix, but approximating the red based on the lightness of the image may not be the right way to go.

Here's the palette:



The left is a power law distribution, the one in the middle is linear, and the one on the right (the one I used) is a combination of the other two.

Anyone else try and replicate an oil painting technique in pixels?

Thanks,
Tourist

[Mathias]

You may've already found it, but I'm reminded of this thread.

[Ai]

I suggest you avoid trying to get a neat 'ramp' and instead accept pairs of similar-brightness, different-hue. Also because of the nature of the source material, and of paintings in general, it doesn't necessarily work very well to pick a set of colors for an individual object but rather I think picking colors for the entire piece is required.
Hence, here is my suggestion for a palette to try:



I approached this as a color-segmentation problem: picking a palette that would produce the most contiguous, readable shapes from the source material after indexization. So I grabbed the pencil tool with a large tip, eyedropped some colors
and laid out sample blobs on top of the painting (in a separate layer). I tried to pick representative colors -- colors that there was relatively a lot of. There's definitely an art to eyedropping from the right place, but once you know it it should be pretty easy to make your own palettes without eyedropping.
I imported the new layer as a palette, decided that the hair colors were needed for coherence and a stronger pink was needed, eyedropped them and added them into the palette.
The result is 16 colors (for the entire painting, not the skin -- there is too much ramp overlap in paintings to
have really distinct ramps)

I believe that in the case of paintings, pure grays and medium-saturation colors are used as utility colors, with high-saturation making up 'spot colors' and pastels comprising the mess of colors that provide more subtle definition.

EDIT:
For working out a decent palette, cutting out the object of interest and indexizing it with a good algorithym (eg. GIMP does a good job) usually gives good results (not necessarily optimal -- most palette optimization algorithyms tend to reduce extreme values, so it's better to ask the paintprogram for a set of say 6 colors which will accurately reflect the gist of the source area, then fill in the inbetween colors yourself. Running Unsharp Mask beforehand with a high radius (16.5 in this case) to counteract the reduction of color extremities is fairly effective.

Miascugh's entry in the thread mathias links seems to get the sort of results you might want.

Hope some of that helps

[Helm]

You don't need so many colors. Actually for the whole painting you need exactly as many colors as Ai has colorpicked above. And this comes down to thinking about your pixel clusters and how they represent the original's shapes and then dithering between them. The chances of making soup instead of drawing are high because well... this is a master painter, but the colors are the easiest part of this.



Here for this very rough pass I'm using the basic techniques that I explained in my pixel cluster text. I make shapes the represent the original with a 10 value palette and then I dither between to represent the blends of the original.

[miascugh]

Miascugh's entry in the thread mathias links seems to get the sort of results you might want.

Here it's noteworthy though, that the skintones of Chun Li are only the slightly modified original Street Fighter ones.

Also, on another note, picking colors for pixel art is much less intuitive than in painting, as you don't layer the different components with their individual reflective properties in pixel art as opposed to some traditional media. You will probably have to (want to) find a compromise between well thought-through color wheel magic to make up for the lack of the "organic" approach to color and effective palettes that are functional in abstracting (compacting) lights behaviour if you're going for something like realism in Pixel Art.

I do find though, that Pixel Art offers some very specific devices that render realism as one of the more boring applications of it . There's a huge, resourceful context to pixels [/hypocrisy]

[Tourist]

Thanks for the feedback all!  I'm off to go practice this for a bit.

Tourist