Synthetic Cubism by Crapadilla

Fantastic work! Reminds me of the works of Russian avant-garde painters of the early XX century. Brilliant.

Quote
Fantastic work! Reminds me of the works of Russian avant-garde painters of the early XX century. Brilliant.

Thank you, Vlad!

Quote
It is slow. You use a lot of Worley noises (Blocks, Cells, Chaffs etc) and you connect them into HighPass -- this is a very slow combo

Aye it is. I have already optimized the worley noises for speed by tuning their details all the way down. The highpasses, I am afraid, are a necessary evil, as they are essential to the filter, creating the dirty collage look as well as accentuating edges and creating depth. I may try and apply the highpass after reassembling RGB and see whether that works well.

Quote
The speed can be improved a bit by eliminating the duplicates of Chaffs components in the bottom-right branch -- these 3 components have the same parameters, therefore they can be replaced with a single Chaffs component with three outbound connections -- that will enable our sample cache optimizations.

Haha, I forgot to randomize these. I will follow your advice instead and turn them into one.

Quote
Another problem is not technical but artistic -- while the filter looks great on our standard preview image (life preserver), it doesn't look as good on other images.

I noticed that the filter works best on "iconic", high-contrast images with well defined areas of light and shadow, clear composition and simpler color schemes with broad areas of color. It certainly won't work for all images. While I am open to suggestions how to improve the generality of the filter, I am still convinced that one won't be able to avoid choosing the subject image very carefully to get best results.

Now that is absolutely impressive, I especially love how only one eye of the puppet gawks from behind the circle on the left side. This truly is art out of one button pressed.

Quote
Crapadilla wrote: I am open to suggestions how to improve the generality of the filter

I'd suggest to work on color -- currently, the filter basically destroys the color scheme of the original image, but I think it can be kept. Look at the HLS and HSY components in Channels -- they can be used to handle Hue, Lightness and Saturation separately.

Quote
Crapadilla wrote: I am afraid, are a necessary evil, as they are essential to the filter

Agreed.

Quote
Crapadilla wrote: Anyway, I would be interested in what exactly you have in mind with the HLS and HSY components. Would you kindly elaborate a bit?

No, I didn't have anything specific in mind, just wanted to suggest the idea of handling color with these components. For example, you can extract Hue and Saturation channels, posterize them with ToneCurve/Stairs combo, then assemble them back to an image -- this can be used to simplify the color scheme.

Quote
Vladimir Golovin wrote: Hmmm. To largely 'destroy' the color scheme of the original actually was my intention.

I've taken a closer look at the way you produce the final color (the Blend component with Color as a blending mode) and it looks fine. Probably I was referring to a specific variation that seemed too discolored to me.

Crapa, I've found the root of all evil -- the elevation gradient is destroying the color. This is because only the topmost row of "pixels" is sampled for a gradient. So, in your situation, the final color scheme depends on what colors show up in the topmost row of the Gradient input. We should find a way to avoid this top-row sampling problem.

Quote
Crapadilla wrote: this would be a problem when the elevation input image doesn't spread its pixel values across the full white to black spectrum, right?

This would be a problem even for a full-range gradient -- if a color is not represented in the top row of the Gradient input, it is not going to appear in the Elevation gradient.

Here's my suggestion -- I've replaced the Elevation Gradient with offset-based distortion, and added a slider to desaturate the image for cases when it is too colorful:

Synthetic Cubism c1u.ffxml

Another kind of image that instantly produces nice results without much tweaking. Obviously the palette sampling is still the problem, as the skin tone is gone.

here goes...

Excellent

Actually, this is a known issue with all gradient inputs -- Filter Forge has no dedicated component and input type for gradients, so we use regular map components and map inputs to emulate them, and this requires certain assumptions that are hard to figure out without documentation.

Yup, really cool results! Congrats Crapa.

- Off-topic -

I noticed that some of the filter pages (including the page for this filter) seem to have a problem with correctly displaying the filter description: Empty spaces between words seem to be randomly omitted.

This may be a problem with my browser (Firefox 1.5.0.4), but you never know. Anyone else notice this?

Quote
Crapadilla wrote: Empty spaces between words seem to be randomly omitted.

Fixed.

A minor problem -- the "Palette View Mode" slider has only 2 positions, it should be replaced with a checkbox.

Just approved the update.

Btw, this is what I get when I render the default preset. I guess the difference is due to 'Seamless Tiling' being activated.

boy, i dont recall seeing this one at all. as steve might say, me likey!