| Summary: | Gaussian Blur filter preview is blocky, the edges of the image are blurred weirdly | ||
|---|---|---|---|
| Product: | [Applications] krita | Reporter: | LukasT <lukast.dev> |
| Component: | Filters | Assignee: | Sahil <nagpal.sahil01> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | canaldin, cberger, dimula73, edu.rm.85, halla, nagpal.sahil01, sven.langkamp |
| Priority: | NOR | ||
| Version: | git master (please specify the git hash!) | ||
| Target Milestone: | --- | ||
| Platform: | Unlisted Binaries | ||
| OS: | Linux | ||
| Latest Commit: | Version Fixed In: | ||
| Sentry Crash Report: | |||
| Attachments: |
Blocks
Blocky artefact |
||
Gaussian Blur filter is broken for a long time. Try using "Blur" filter. It works fine. But, afair, it doesn't work without fftw3 library. Can you check if this still happens? Just for the record, before everyone has gone to bed and forgets about it: 21:07:36 < slangkamp> why is the filter src paint device exactBounds limited to the rect that the thread is working on? 21:07:56 < boud> it is? 21:08:11 < boud> a filter should be able to read from the whole src device iirc 21:09:25 < slangkamp> I run the gausian blur filter and from applyMatrix I get src->exactBounds() I get lots of different values 21:09:55 < slangkamp> which appear to be the areas the threads are working on 21:10:08 < boud> maybe because now the filters copy the bit the thread works on onto something temporary? 21:10:59 < boud> but I've lost track of that code a bit 21:11:57 < slangkamp> the filter needs the complete bounds otherwise the repeat iterators are repeating at the wrong bounds 21:12:41 < boud> is this in the mask or mask preview, or when apply and close the filter dialog ? 21:16:20 < slangkamp> let me check 21:17:11 < slangkamp> only with preview 21:17:30 < boud> and if you create a filter mask? 21:20:42 < boud> slangkamp: but for this issue of yours, we really need dmitry. It's tied together with transactions, changerects, needrects and things 21:20:54 < slangkamp> hmm there is something else going wrong 21:21:14 < slangkamp> the preview takes several times longer than the apply 21:22:08 < boud> that's probably because the preview is done in chunks, with an initialisation cost for every chunk. Best bet is to profile it. 21:26:30 < slangkamp> boud: calculated too often http://pastebin.com/DJ4Ax9wq image is 1024x768 21:26:59 * slangkamp is away from 30 minutes 21:27:50 < boud> slangkamp: very strange 21:31:00 < dmitryK41> boud, slangkamp: yes the merger creates a temporary area to work *if needed* 21:31:24 < boud> dmitryK41: what determines the need? 21:31:35 < dmitryK41> boud, slangkamp: so the filter must be able to work locally 21:32:14 < boud> but that's always going to be a problem with convolution filters -- they need to read beyond the bounds of local areas. 21:32:21 < dmitryK41> boud: e.g. when needRect!=requestedRect!=changeRect 21:32:48 < dmitryK41> boud: this problem has been solved with needRect mechanism 21:33:16 < dmitryK41> boud: now the merger prepares some excessive area for the filter to work 21:33:57 < boud> dmitryK41: https://bugs.kde.org/show_bug.cgi?id=245870 is still a problem, and also slangkamp found that the same area seems to be calculated a lot: 21:34:03 < bugbot> KDE bug 245870 in krita (general) "Gaussian Blur filter preview is blocky, the edges of the image are blurred weirdly" [Major,New] 21:34:09 < boud> 21:26:30 < slangkamp> boud: calculated too often http://pastebin.com/DJ4Ax9wq image is 1024x768 21:34:38 < dmitryK41> boud: and if the convolution filter is in the middle of the masks stack you simply can't apply it on the layer directly 21:34:59 < boud> yes, I understand that 21:35:09 < boud> but we've still got a problem :-) 21:36:10 < dmitryK41> boud: as far as i remember Gaussian Blur filter was broken itself 21:36:23 < dmitryK41> boud: that is why we had these blocky results 21:36:36 < dmitryK41> boud: you can check Blur filter 21:36:44 < dmitryK41> boud: it works fine 21:37:04 < dmitryK41> boud: and it is convolution filter as well 21:37:27 < dmitryK41> boud: Blur != Gaussian Blur 21:38:30 < boud> best wait a bit -- slangkamp should be back in twenty minutes and he has more details. Created attachment 52653 [details]
Blocky artefact
Here is showed the blocky artefact. Draw some lines throught the whole canvas and then use gaussian blur. The place of the blocky effect is both in the preview and then in the result.
The edges are nicer now.
Here is the log from valgrind http://pastebin.ca/1966873 (temporary available) After running Krita in valgrind, the bug is not reproducible. Mighty valgrind healed the bug for me. SVN commit 1187822 by lukast: Fix for the border artefacts Proposed by Cyrille Berger CCBUG:245870 M +5 -1 kis_gaussian_blur_filter.cpp WebSVN link: http://websvn.kde.org/?view=rev&revision=1187822 It is not blocker anymore, the implementation of BORDER_WRAP (ie the behaviour near the edge of the image) is actually a new feature. And should probably be reported as a different bug report for clarity. Ok, I will report new bug report :) Closing as fixed then. commit 2eece610616003b494a84fcb4ae7ff084c35ab0f branch words-change_tracking-ganeshp Merge: 4879e17 be16917 Author: Ganesh Paramasivam <ganesh@crystalfab.com> Date: Thu Oct 21 08:18:21 2010 +0000 Merged revisions 1187683,1187692,1187696,1187699,1187724,1187736,1187740-1187741,1187749,1187768-1187770,1187778,1187799-1187800,1187815,1187817,1187822,1187833,1187839,1187868,1187913,1187917,1187954,1187967,1187974 via svnmerge from svn+ssh://ganeshp@svn.kde.org/home/kde/trunk/koffice ........ r1187683 | vandenoever | 2010-10-20 11:48:05 +0530 (Wed, 20 Oct 2010) | 1 line Catch errors resulting from XML that is not wellformed. ........ r1187692 | vandenoever | 2010-10-20 12:59:45 +0530 (Wed, 20 Oct 2010) | 3 lines Do not write out text:page-sequence. It is only allowed when no text elements like text:p or text:h are written. KOffice does not seem to use text:page-sequence at all. ........ r1187696 | vandenoever | 2010-10-20 13:39:24 +0530 (Wed, 20 Oct 2010) | 3 lines move style:master-page-name and style:master-page-name KOffice was writing style:master-page-name and style:master-page-name attributes as part of style:paragraph-properties instead of styles:style. This is fixed now. ........ r1187699 | vandenoever | 2010-10-20 13:45:47 +0530 (Wed, 20 Oct 2010) | 2 lines Do not write attribute koffice:name. It is not used and invalid. ........ r1187724 | vandenoever | 2010-10-20 15:53:18 +0530 (Wed, 20 Oct 2010) | 1 line Add script to validate ODF. ........ r1187736 | breitmeyer | 2010-10-20 16:31:34 +0530 (Wed, 20 Oct 2010) | 2 lines loaded axis minimum and maximum values ........ r1187740 | lassin | 2010-10-20 16:52:09 +0530 (Wed, 20 Oct 2010) | 2 lines Fixed drawingML bullet color reading, patch provided by Nandita. ........ r1187741 | vandenoever | 2010-10-20 16:54:51 +0530 (Wed, 20 Oct 2010) | 3 lines Save koffice:name (which is not valid ODF!) but only if it is absolutely necessary. That is, if it is not empty and not the same as the 'official' draw:name value. For future KOffice versions it might be better to store this information in an RDF graph, so that the content.xml can be valid ODF. ........ r1187749 | lassin | 2010-10-20 17:05:47 +0530 (Wed, 20 Oct 2010) | 2 lines Made symbol bullets in ooxml to behave sanely in more cases. ........ r1187768 | mlaurent | 2010-10-20 17:22:48 +0530 (Wed, 20 Oct 2010) | 2 lines Install as program ........ r1187769 | mlaurent | 2010-10-20 17:23:03 +0530 (Wed, 20 Oct 2010) | 2 lines install as program ........ r1187770 | mlaurent | 2010-10-20 17:23:16 +0530 (Wed, 20 Oct 2010) | 2 lines Install as progrzm ........ r1187778 | lassin | 2010-10-20 17:39:25 +0530 (Wed, 20 Oct 2010) | 2 lines Simplified drawingML & pptx code. ........ r1187799 | vandenoever | 2010-10-20 18:31:59 +0530 (Wed, 20 Oct 2010) | 1 line Add validation of manifest.xml ........ r1187800 | vandenoever | 2010-10-20 18:32:03 +0530 (Wed, 20 Oct 2010) | 1 line Add required manifest version number. ........ r1187815 | lassin | 2010-10-20 18:57:52 +0530 (Wed, 20 Oct 2010) | 2 lines Fixed group shapes in drawingML to have a draw style. ........ r1187817 | lassin | 2010-10-20 18:58:43 +0530 (Wed, 20 Oct 2010) | 2 lines Simplified pptx code. ........ r1187822 | lukast | 2010-10-20 19:12:56 +0530 (Wed, 20 Oct 2010) | 5 lines Fix for the border artefacts Proposed by Cyrille Berger CCBUG:245870 ........ r1187833 | vandenoever | 2010-10-20 19:49:45 +0530 (Wed, 20 Oct 2010) | 1 line Handle KeyError that is thrown if a file is missing from the document. ........ r1187839 | vandenoever | 2010-10-20 19:51:17 +0530 (Wed, 20 Oct 2010) | 1 line Fix path to manifest.xml. ........ r1187868 | breitmeyer | 2010-10-20 21:04:47 +0530 (Wed, 20 Oct 2010) | 2 lines Fixed loading of cell regions that contain lists of regions ........ r1187913 | staniek | 2010-10-21 00:11:41 +0530 (Thu, 21 Oct 2010) | 2 lines SVN_SILENT don't use reference ........ r1187917 | staniek | 2010-10-21 00:23:24 +0530 (Thu, 21 Oct 2010) | 4 lines KexiDB *add sanity checks to assignment operators ........ r1187954 | staniek | 2010-10-21 02:32:21 +0530 (Thu, 21 Oct 2010) | 6 lines Correction for r1183552: "Make KoReport a full koffice-libs member" For TINY target reports should be switched off, as well as report plugins, so the change in r1183552 now applies only to the default (desktop) builds. ........ r1187967 | langkamp | 2010-10-21 03:02:44 +0530 (Thu, 21 Oct 2010) | 1 line simplified code ........ r1187974 | langkamp | 2010-10-21 04:08:53 +0530 (Thu, 21 Oct 2010) | 3 lines update specific color selector colorspace when image colorspace is changed CCBUG:252461 ........ svn path=/branches/work/koffice-change-tracking/; revision=1188051 Bug still does appear after the filter is applied. Hello, I get the border's black alpha bleeding into the canvas on all sides except the left one. reference here: http://www.mediafire.com/?vem5vjb7c7yql5v Thank you I'm using Krita 2.4 on ubuntu 12.04 on a Lenovo w520 (linux certified) Seems related to 296821 The thing is, this did work correctly at one point -- I wonder what the issue here is. Alexis -- you are on *buntu or windows? ubuntu as stated *** Bug 296821 has been marked as a duplicate of this bug. *** Bug is now fixed in 'krita-testing-kazakov' branch, please test and report |
Created attachment 49513 [details] Blocks Version: svn trunk (using KDE 4.4.5) OS: Linux Gaussian blur is little less useful then the one in GIMP when I post-process my painting. First problem is that the preview shows some big rectangular blocks on the image. See attachment. Second problem is that when bluring, the image is repeated in invisible area. It is good for some filters, but in this case it looks bad. Reproducible: Always Steps to Reproduce: 1. Run Krita 2. Paint some masterpiece or use gradient to fill the canvas to the edges 3. Filter-> Gaussian Blur 4. Preview is blocky 5. Apply the filter, result is repeated and looks weird Actual Results: The image looks strange when trying to post-process the painting. I would propose to copy behaviour from GIMP in this case. OS: Linux (i686) release 2.6.32.9-70.fc12.i686 Compiler: gcc