rust-fractal 0.7.2

  • 34 Replies
  • 1010 Views

0 Members and 1 Guest are viewing this topic.


Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #16 on: October 05, 2020, 05:54:55 AM »
I'm just about to release a new update which should address the memory usage problems - this is done by introducing a new option in the config files which defines how often the SA and reference data is stored. For the 'seahorse tiling' image I set this to 1000 and the program used a maximum of about 3.2 GB of RAM but took nearly an hour (max iters was more than 45 million).

I've now put the enhanced probe points implementation into the stable algorithm. I encountered no issues with this. However, the tiled SA does not work 100% of the time and is left in the experimental options. Most images work but it seems that some probe points still cause overskipping problems which will need to be investigated further.

Update: 0.10.0 released https://github.com/jackyarndley/rust-fractal/releases/tag/0.10.0
« Last Edit: October 05, 2020, 07:31:36 AM by unassigned »

Offline gerson

  • *
  • Fractal Frogurt
  • ******
  • Posts: 432
« Reply #17 on: October 10, 2020, 08:27:08 PM »
Testing some palettes with new version.

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2244
« Reply #18 on: October 10, 2020, 08:44:27 PM »
Testing some palettes with new version.
Looks like something is wrong with your aspect ratio.

Offline gerson

  • *
  • Fractal Frogurt
  • ******
  • Posts: 432
« Reply #19 on: October 10, 2020, 08:50:11 PM »
It is a 3600x2400 resized to 1600x1200. Now it is proportional.
« Last Edit: October 11, 2020, 03:44:40 AM by gerson »

Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #20 on: October 17, 2020, 02:30:04 AM »
I've found a bug in the tiled series approximation for the experimental options - fixing this and I haven't found any issues. It's much faster on images which have quite a lot of iteration depth to them, in some cases I have seen images render up to 3x faster. In my test suite there is approximately a 40% speedup!

Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #21 on: October 17, 2020, 02:56:13 AM »
rust-fractal 0.11.0 is now out with the fixes to the tiled series approximation. It can be downloaded here:

https://github.com/jackyarndley/rust-fractal/releases/tag/0.11.0

I would really like for the tiled series approximation to be tested. It is enabled using the 'experimental = true' flag in the config files, and is on by default using the high.toml options file. If anyone has some particularly difficult locations please let me know and I can test them. It offers some substantial speedups to the iteration stage of rendering; for the 'wiggly fungus spores' location, iteration went down from 20s to 6s.

Offline gerson

  • *
  • Fractal Frogurt
  • ******
  • Posts: 432
« Reply #22 on: October 18, 2020, 07:37:15 PM »
This days I was rendering yours example's images. Only 5 (e52464, opus2, orion, seahorse and super_dense) I didn't get because memory alocation. Will try in another computer.
Was trying new palettes and liked a lot.
If possible, to do a better documentation (help) to how to use it.
And a suggestion to save image with an automatic name as rust-date-time-name
like Infinity factal does ( InfinityFractal-Mandelbrot-20201014-194702.ifp)
it would be like this: rust-20201014-194702_00000000_3.70E191

Here your 11_dimensions.toml with a grayscale palette.

Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #23 on: October 20, 2020, 05:50:58 AM »
I will try and improve the documentation soon - its on my list of things to do. With respect to your suggestion it sounds like a good idea but I need for the keyframe number to be at the start so that they can easily be ordered. Maybe something like 00000000_3.70E191_RUSTFRACTAL or something similar would be appropriate - the date and time of creation is already in the file metadata but I do realize it is sometimes useful in the file name.

I've been working on implementing analytic derivative calculation to rust-fractal. Currently it is working well and should be in the next update once I've done a little more testing and got the EXR export for the directional derivative working properly. In some tests it takes approximately 1.2 to 2.0 times the time to calculate the derivatives in addition to the iteration counts.

Below is 'wiggly fungus spores' location, rendered with DE in 1920x1080. It takes my computer (i7-4790k) 45 seconds to compute.
« Last Edit: October 20, 2020, 06:02:06 AM by unassigned »

Offline claude

  • *
  • 3f
  • ******
  • Posts: 1622
    • mathr.co.uk
« Reply #24 on: October 20, 2020, 11:40:44 AM »
Below is 'wiggly fungus spores' location, rendered with DE in 1920x1080.
Looking good!

Being very picky, there is a bit of Moire artifacting near the edges of the main features that could be traded for noise by using jitter (picking a uniformly pseudorandom location in the pixel's region instead of the pixel center, with the offset different for each pixel).  Personally I find noise much less visually distracting than Moire.

Offline gerson

  • *
  • Fractal Frogurt
  • ******
  • Posts: 432
« Reply #25 on: October 20, 2020, 09:48:00 PM »
About documentation is due many of us are just user, not programmer, so will help us to know how to use it beter.
For example, what occurs when change approximation_order or glitch_percentage...
I notice that low and high.toml have rotate inside, and some locations have too (like x.toml). Is the same or what that counts?
I rendered x.toml with a little more zoom (zoom = "3.5E236") and custom palette, see if it is ok. Look at the center of espiral.

Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #26 on: October 20, 2020, 11:14:54 PM »
I'll definitely get some better documentation done soon.

glitch_percentage is the maximum percentage of the image which is left 'glitched' with respect to glitches during perturbation. If you want to see where these glitches are set the option 'display_glitches' to true and turn the glitch percentage to something like 99.0 so that the glitch correction algorithm will not run.

approximation_order is the order of the polynomial used for series approximation. Setting this to higher numbers will generally increase the number of skipped iterations but increase the chances that there are errors in the image. Generally, for a large image 64 works well, any more there starts to be more and more errors depending on the image.

I've rendered that image using DE and the center seems to be there for me. Did you use the high.toml file used in the latest release? To me it looks like the glitch correction hasn't worked or not run enough, maybe try running the image with the same settings and the display_glitches option turned on and look for red regions.

Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #27 on: October 20, 2020, 11:46:19 PM »
Being very picky, there is a bit of Moire artifacting near the edges of the main features that could be traded for noise by using jitter (picking a uniformly pseudorandom location in the pixel's region instead of the pixel center, with the offset different for each pixel).  Personally I find noise much less visually distracting than Moire.

How much jitter do you think it appropriate? I've done some tests with [-0.2,0.2] pixel spacing jitter and it seems to improve it a bit, but with more it just gets really bad especially with some of the fine filament details. Attached is an image with this applied.

Offline gerrit

  • *
  • 3f
  • ******
  • Posts: 2244
« Reply #28 on: October 21, 2020, 04:00:14 AM »
In bad cases I oversample factor 10, with jitter (random on full pixel square, not sure what your 0.2 means) , then Gaussian blur radius 5 pixels, then downsample.
PS In really hard cases render many times as above and average before doing all that.

Offline unassigned

  • *
  • Fractal Fanatic
  • ***
  • Posts: 35
« Reply #29 on: October 21, 2020, 05:43:19 AM »
Julia morph (e22522) with analytic DE working. Really shows the internal structure very clearly.
found another error which gives convincing images... fixed
« Last Edit: October 21, 2020, 07:34:47 AM by unassigned »