lycium
|
|
« Reply #135 on: March 20, 2016, 05:49:22 AM » |
|
it seems this is the last big hurdle that needs to be properly solved for the series approximation / perturbation technique to be considered good to go / non-experimental. so far it seems we've all been more or less silently agreeing to lie to ourselves a bit and pretend that this is basically already good to go, though of course in reality if something only works "maybe" or "some of the time" or "to a certain degree" or "its anybody's guess," it is not in fact good to go.
I 100% agree, and have been following the SFT / perturbation thing since the beginning, spoke with the author of UltraFractal about his investigations of the technique in 2014 at the Fractal Art Symposium... my comment from 2013 still stands: "Nevertheless, the real issue I have with this method is that it's not a true acceleration for the same computation. It's a kind of approximation, and the deviation from the "true" result is essentially impossible to quantify since it's a chaotic system. So it's something like apples and oranges comparison, no?" I don't see how trading computer processing power for human effort (looking for glitches) is a winning strategy, or even a sure one - people have missed glitches before, and until there's some kind of proof of correctness (itself more difficult to compute than a reference image with direct iteration) the resulting image quality is dependent on the patience and eyesight of the person doing the rendering.
|
|
|
Logged
|
|
|
|
stardust4ever
Fractal Bachius
Posts: 513
|
|
« Reply #136 on: March 20, 2016, 09:11:45 AM » |
|
I am almost tempted to do something. We know there is a donut shaped zone of iteration data with a guaranteed failure every time when the reference pixel is located within it. If I zoom directly into this "no man's land" immediately outside the target formation, pick an inflection point, and continue the zoom until 50% deeper into the set where the fractal data essentially doubles itself, will the formations on either side of the new centroid be messed up? This will be a fun experiment to try later this week when I have some free time. I don't see how trading computer processing power for human effort (looking for glitches) is a winning strategy, or even a sure one - people have missed glitches before, and until there's some kind of proof of correctness (itself more difficult to compute than a reference image with direct iteration) the resulting image quality is dependent on the patience and eyesight of the person doing the rendering.
I wouldn't say that exactly. Most of the Mandelbrot glitches I have seen are quite glaring. The Mandel Machine glitches often look like "deflated beach balls" and the Kalles Fraktaler glitch frames, when they occur in a movie, are typically the result of a forked path that travels away from the centroid. My objective is to find an image with the reference point focused about the true centroid relative to the rendered pattern, and still glitch out, making the image and the area surrounding it truly unsolvable through series approximation. A zoom path that deliberately forks at a location within the "donut of glitches" seems like a sure fire way to test the system.
|
|
« Last Edit: March 20, 2016, 09:20:35 AM by stardust4ever »
|
Logged
|
|
|
|
hapf
Fractal Lover
Posts: 219
|
|
« Reply #137 on: March 20, 2016, 09:51:31 AM » |
|
I 100% agree, and have been following the SFT / perturbation thing since the beginning, spoke with the author of UltraFractal about his investigations of the technique in 2014 at the Fractal Art Symposium... my comment from 2013 still stands: "Nevertheless, the real issue I have with this method is that it's not a true acceleration for the same computation. It's a kind of approximation, and the deviation from the "true" result is essentially impossible to quantify since it's a chaotic system. So it's something like apples and oranges comparison, no?" There are safe ways to do this. They are simply not as fast as the less safe ways. But still magnitudes faster than the full arbitrary precision version when dealing with deeeep zooms. One should never forget that the perturbation approach unlike series approximation is no approximation but 100% correct like the original iterations. The glitches are due to accumulation of rounding errors, not an approximation in the formula. So there is always the possibility to reduce rounding errors by using more bits for the deltas. While this makes things a lot slower again the bits needed are far less than needed for the original iterations at full precision when zooming very deep. So speedups are still big. And series approximation can be made very safe if one is careful and not too ambitious with the percentage of skipped iterations. The problem there is that we lack a clear theoretical criterion that tells us it's safe till here and not beyond.
|
|
« Last Edit: March 20, 2016, 10:20:29 AM by hapf »
|
Logged
|
|
|
|
hapf
Fractal Lover
Posts: 219
|
|
« Reply #138 on: March 20, 2016, 10:22:09 AM » |
|
A zoom path that deliberately forks at a location within the "donut of glitches" seems like a sure fire way to test the system. If you give me such a location I can try to see what is causing this.
|
|
|
Logged
|
|
|
|
|
quaz0r
Fractal Molossus
Posts: 652
|
|
« Reply #140 on: March 20, 2016, 07:22:18 PM » |
|
There are safe ways to do this ... The problem there is that we lack a clear theoretical criterion that tells us it's safe till here and not beyond. my whole point was exactly this, regarding the series approximation. someone has yet to formulate a proper criterion which ensures the whole thing wont blow up, let alone getting you as far as you can possibly get with it in addition to ensuring its safety. again i would assert that implying it is basically good to go, or making allusions to its theoretical safety, does not equate to real-world correctness. it looks like we all have implementations that have been shown to blow up and produce incorrect results. not good to go.
|
|
|
Logged
|
|
|
|
quaz0r
Fractal Molossus
Posts: 652
|
|
« Reply #141 on: March 20, 2016, 07:57:38 PM » |
|
with regard to the approach botond described, checking that each term is smaller by some amount, maybe this amount should be dynamically set based on the zoom depth and/or some other criteria? are any of you already doing something like that? it seems that smaller values work at lower depths, with higher depths indicating larger values?
|
|
|
Logged
|
|
|
|
hapf
Fractal Lover
Posts: 219
|
|
« Reply #142 on: March 20, 2016, 08:44:49 PM » |
|
I tried to reproduce the effect with overskipping with references in the region but could not. The errors look different and when skipping more (beyond the minimal iterations in the image) the outer parts become featureless, but the center does not become a donut. So this looks like a bug and not simply overskipping. How much do the programs want to skip here? What is the iteration max of the reference?
|
|
|
Logged
|
|
|
|
hapf
Fractal Lover
Posts: 219
|
|
« Reply #143 on: March 20, 2016, 09:00:34 PM » |
|
with regard to the approach botond described, checking that each term is smaller by some amount, maybe this amount should be dynamically set based on the zoom depth and/or some other criteria? are any of you already doing something like that? it seems that smaller values work at lower depths, with higher depths indicating larger values?
I tested many variants of this and with fewer coefficients it seems to be rather safe when wanting to skip with x coefficients to make sure all contributions from another x coefficients after that individually and together compared to the first x is so small that the absolute exponent difference is >= the negative exponent of the pixel size. But with more and more coefficients (128...) even this had cases that had errors above 1 percent, or 5. The more coefficients the more the contributions are spread over more and more coefficients going up and down in waves. The waves jump from level to level (based on the "keys to the kingdom") but the overall effect is hard to establish when coming closer to the maximum skip possible and many coefficients used. When playing it safe I simply could not get close to the real optimal skipping possible found with actually testing the image at various locations, at least not with many coefficients and all locations I looked at.
|
|
|
Logged
|
|
|
|
stardust4ever
Fractal Bachius
Posts: 513
|
|
« Reply #144 on: March 21, 2016, 11:06:07 AM » |
|
I tried to reproduce the effect with overskipping with references in the region but could not. The errors look different and when skipping more (beyond the minimal iterations in the image) the outer parts become featureless, but the center does not become a donut. So this looks like a bug and not simply overskipping. How much do the programs want to skip here? What is the iteration max of the reference?
The "donut" in this instance is not a feature persay but a narrow band of iteration data surrounding the feature that produces glitches when the reference pixel is contained therein. Reference pixels within this donut shaped zone produce glitches using default settings in Mandel Machine v1.3.15.
|
|
« Last Edit: March 22, 2016, 12:50:39 AM by stardust4ever »
|
Logged
|
|
|
|
quaz0r
Fractal Molossus
Posts: 652
|
|
« Reply #145 on: March 21, 2016, 07:46:22 PM » |
|
mandel machine might be set too lax in its glitch detection/correction in addition to the overskipping issue
|
|
|
Logged
|
|
|
|
stardust4ever
Fractal Bachius
Posts: 513
|
|
« Reply #146 on: March 22, 2016, 01:33:44 AM » |
|
@Kalles Fraktaler: Kalles Fraktaler is failing to find the centroid again. I am using latest version. Size = 640x360 Zoom = 16 "Find Minibrot" command fails by zooming into the chromosome instead of between them. I have zoomed well into the glitch area of the mmf I posted on previous page. I zoomed up until this "64X" location using Mandel Machine, then decided to automate the process, so entered the coordinates into KF and used your "find minibrot" menu item and it fails. I assume it would repeatedly fail again and again to locate the centroid once it reaches a sufficiently long chromosome formation, and this error would likely repeat ad infinitum. I wanted to find the periodic double of a known glitch location from within the Mandel Machine file I posted earlier but didn't feel like manually zooming to the approximately 2^-5535 zoom depth this would require. I only want to test the location to see if Mandel Machine barfs on it. EDIT: I found the position manually. Apparently, this location I searched for doesn't glitch up at all. To view the original "glitch" you must reset the magnification to 3470. MMF parameters... image.width=1280 image.height=720 image.supersampling=2 position.re=-1.7697392710757027841152006141906594654234779937259515247210119204006409471133661312119570293427350953111092082771973598855782759599980633037132375185987670114621045880606454417706999699982839165518556104246980851140230022353588311445547248451142336585544111916489341441418890633019730101854648463669109636471162672120817824917407218303348419898985242808140058964414457471465221536301689319210853379995524156624829942780946758964309072692484524503974031949154005423948971844113769257156823395255042997434557967120122936287886272994336827603988076644070845456535059390110455532123959338354686709501965947157116618024249418340042125809823163860665551942191883280150556503637746217516167947991343226606484863461721881250950867469305763323991190602711636193429181021868663451423240656141036481686672207216510178940927578426246138789051935948202201039069515784638090342979600881257471340705419667858851219685966190797231145977198268039243169238831169502470679242625047521216680080745102102338697978848434690710623615151918273912523550802047059388477610120227482645505202901268310259531375914654127989863575658517285857298644046776833979372298401170515594247650694236040835088580636725197188883301838633826884891907725215420267621408231407868087863857155606555607233691975269649870535381564546102514139884292840878611927606673694598379606020367361658483137883367223105802511967374846046867768244164444838768374265954854033947228261298148439975283873913669268210784821270043045378977203278631694546469555524912082733758594148241573505499833138090100150440411001854828832965244988 position.im=0.0047957559338652178505615663220486971154304886886673920781256111752398106107280366675989298691779797136737364961830758523741254480286664920460479195128211953261155569223186412660856033539583067793744552698972305956559820659036789189846739491155635838617923497250059314208186984179036223826655324004181421335633738898275114135499319014247271095039118772708132302750547913645230844806438628351797278101552420985568133578867374426937467785916207019758747794073870494765108717384461786347756065294984135627296206900076429271573756624100323471123560470223626746460265018297930359289043411279245762239698772188260974285993117734181827318766613834051977484272020698249213489627750957498023989747438133307647274695802671234943834981584342636056200655633883556867650337796993091229444120358050723663523120235639704136824323264324209617146569763642761341662189958962136924719552069482293150925374378282906250039171082284465702486444460840842205856092252444604814892208734287500540251296138686660257059365390928468879124177736053405984117806338296754892837341507069125502543659126775790197583562065309265743403930782970442835855756964862590861406611214408287544430346102387396417375654652294552820157169020238943731165207653089997593275027403759259891988184570391468326105306257559441098144027014942084582917715350747076407623299327166201532840215419753271633485048688749655561442706732391867989832423455475081605888333753784925707440085445028189173333773877469260631083482242534267212635500091888677779450110195272426094961657170920294279611308694879151344530550580595486867216094 position.magnification=5197.0 position.rotation=90.0 computation.iteration_limit=1000000 rendering.inner_color=000000 rendering.outer_color=dddd00 rendering.transfer_function=1 rendering.color_density=-142.000 rendering.dwell_bands=0 rendering.palette=0.0,85,0,85,0.1,170,0,0,0.2,255,85,0,0.3,255,170,0,0.4,255,255,85,0.5,170,255,170,0.6,85,255,255,0.7,0,170,255,0.8,0,85,255,0.9,0,0,170 rendering.color_offset=501760 rendering.bump_mapping=false rendering.bm_light_direction=0 rendering.bm_depth=50 rendering.bm_strength=50
|
|
« Last Edit: March 22, 2016, 03:10:47 AM by stardust4ever »
|
Logged
|
|
|
|
stardust4ever
Fractal Bachius
Posts: 513
|
|
« Reply #147 on: March 22, 2016, 03:08:13 AM » |
|
The above coordinates at 3470 zooms: Shown below are Automatic Precision and Float Precision. All of the other available scaled precision levels selectable within Mandel Machine produce a glitch comparable to either of the two below images, so this exact coordinate is not-correctable without changing the reference point or beefing up the precision somehow.
|
|
|
Logged
|
|
|
|
stardust4ever
Fractal Bachius
Posts: 513
|
|
« Reply #148 on: March 22, 2016, 03:19:07 AM » |
|
Last one, I promise, pinky swear... You can have fun deliberately perturbimg the image by forcing the wrong settings! X of Xs (automatic) vs X of Xs (floating point) "Fun With Glitches.mmf" image.width=1600 image.height=1600 image.supersampling=2 position.re=-1.7697392710757027841152006141906594654234779937259515247210119204006409471133661312119570293427350953111092082771973598855782759599980633037132375185987670114621045880606454417706999699982839165518556104246980851140230022353588311445547248451142336585544111916489341441418890633019730101854648463669109636471162672120817824917407218303348419898985242808140058964414457471465221536301689319210853379995524156624829942780946758964309072692484524503974031949154005423948971844113769257156823395255042997434557967120122936287886272994336827603988076644070845456535059390110455532123959338354686709501965947157116618024249418340042125809823163860665551942191883280150556503637746217516167947991343226606484863461721881250950867469305763323991190602711636193429181021868663451423240656141036481686672207216510178940927578426246138789051935948202201039069515784638090342979600881257471340705419667858851219685966190797231145977198268039243169238831169502470679242625047521216680080745102102338697978848434690710623615151918273912523550802047059388477608402978 position.im=0.0047957559338652178505615663220486971154304886886673920781256111752398106107280366675989298691779797136737364961830758523741254480286664920460479195128211953261155569223186412660856033539583067793744552698972305956559820659036789189846739491155635838617923497250059314208186984179036223826655324004181421335633738898275114135499319014247271095039118772708132302750547913645230844806438628351797278101552420985568133578867374426937467785916207019758747794073870494765108717384461786347756065294984135627296206900076429271573756624100323471123560470223626746460265018297930359289043411279245762239698772188260974285993117734181827318766613834051977484272020698249213489627750957498023989747438133307647274695802671234943834981584342636056200655633883556867650337796993091229444120358050723663523120235639704136824323264324209617146569763642761341662189958962136924719552069482293150925374378282906250039171082284465702486444460840842205856092252444604814892208734287500540251296138686660257059365390928468879124177736053405984117806338296754892838258823 position.magnification=3473.0 position.rotation=-8.5 computation.iteration_limit=400000 rendering.inner_color=000000 rendering.outer_color=dddd00 rendering.transfer_function=1 rendering.color_density=-142.000 rendering.dwell_bands=0 rendering.palette=0.0,85,0,85,0.1,170,0,0,0.2,255,85,0,0.3,255,170,0,0.4,255,255,85,0.5,170,255,170,0.6,85,255,255,0.7,0,170,255,0.8,0,85,255,0.9,0,0,170 rendering.color_offset=221184 rendering.bump_mapping=false rendering.bm_light_direction=0 rendering.bm_depth=50 rendering.bm_strength=50
Set scale precision to "float" and deselect "autocorrect glitches" Animated version: http://sta.sh/01xwujdi80tx
|
|
« Last Edit: March 22, 2016, 04:12:43 AM by stardust4ever »
|
Logged
|
|
|
|
PieMan597
|
|
« Reply #149 on: March 22, 2016, 10:21:01 AM » |
|
Maybe unchecked "ignore small addends?"
|
|
|
Logged
|
|
|
|
|