Title: large render test Post by: quaz0r on February 20, 2015, 12:45:57 AM large render test
(http://nocache-nocookies.digitalgott.com/gallery/17/9373_20_02_15_12_45_56.png) http://www.fractalforums.com/index.php?action=gallery;sa=view;id=17068 just a simple image, my first test of the tiled rendering mode i added to my mandelbrot renderer for achieving larger resolutions and more supersampling with a lower memory footprint. with 10x supersampling, effectively a 90000x60000 render filtered and downsampled to 9000x6000. no slope effect on this one, next i need to fix slope effect for tiled rendering mode. full image is 21MB: http://filebin.net/cs9sv2gdpp/tiletest1.png Title: Re: large render test Post by: claude on February 22, 2015, 09:51:41 PM been thinking about fixing tiled rendering for mightymandel too. one strategy might be to render tiles with a slightly larger region (as edges get corrupted with slope encoding / fake-de), then crop them to the desired size. if the tiles are NxN, you only need around k*N extra edge samples, where k is typically a lot smaller than N. so it computes more iterations, but assembling later is a lot easier (no postprocessing of combined tiles needed).
btw, probably my monitor is not set up correctly, but I noticed some visible banding in the dark areas (I amplified a small region of the attached) - maybe dithering would help? ie, compute the "perfect" supersampled colour for each pixel as an rgb floating point triple in [0..1], then add a small amount of noise (peak-to-peak amplitude 1/256 for 8bit final image) before rounding to rgb int8 triple in [0..255]. dithering reduces perceptual quantization error, like banding at low amplitude changes - the steps between bands add visually jarring high frequency content, dithering spreads out the error spectrum. same theory applies to digital audio. I'm sure wikipedia or some dsp site could explain it better than i can... i'll start a new thread once i've researched it a bit more... Title: Re: large render test Post by: quaz0r on February 22, 2015, 10:46:42 PM yeah, i was thinking the same thing about slope. as far as the banding, i know there should be some way of correcting it, im just not sure how to go about it. i was thinking i probably dont want to just apply some generic filter to the whole image? but as far as a more surgical strike im not sure what form that should take exactly. i guess i figure it should have to do with adding more magic to the smooth iterations math, which itself is just another formula someone else came up with that i dont really understand :angry: Quote i'll start a new thread once i've researched it a bit more... if you came up with more thoughts on the matter that would be great :D |