Hi, I'm the one who posted that question on Stack Overflow.

I've tested this code, and it doesn't fix the underlying problem -- the "cleared" (alpha=0) contents of the renderer are still being blended into the desired background color; the problem is merely masked with the included test because the test background and the cleared renderer are both black.

I did some Googling and some digging into AGG (notably, findingthis) and I've found the culprit: It's the use ofagg::pixfmt_rgba32 instead ofagg::pixfmt_rgba32_plain as the pixel format for the renderer. The blending function for the non-plain version takes a shortcut in the calculation that breaks down when the existing value in the buffer has alpha=0; the_plain variant handles this correctly. I've done a quick test of a build that uses_plain instead, and the output it generates is correct.

Should I submit a pull request with that fix, and some improvements to the test case?

Sure. You can submit them as PRs against my PR, or its own PR to mpl.

Cheers@Westacular

Thanks for the work on this. We've found a lot of subtle things out about Agg lately, haven't we?

@pelson: I wonder whether if maybe we should rebase this on v1.2.x, since it's really sort of bugfix. I guess I'm waffling -- because obviously it will change behavior who are expecting transparent to come out as white in the PNG. There have been a few other notable bugfixes on v1.2.x since 1.2.1, and I'm wondering if it might be nice to have a 1.2.2 sooner rather than later.

We've found a lot of subtle things out about Agg lately, haven't we?

Yes. As we speak I'm going through the Agg backend to have another look at compound rendering, so I might turn over some more too...

There have been a few other notable bugfixes on v1.2.x since 1.2.1, and I'm wondering if it might be nice to have a 1.2.2 sooner rather than later.

I certainly have a couple of bugs that need my attention before a v1.2.2, but I also note that v1.3 athttps://github.com/matplotlib/matplotlib/blob/master/doc/users/whats_new.rst#new-in-matplotlib-1-3 lists more new features than any other release. There is definitely a debate to be had on the dev mailing list about when to start aiming for a v1.3 (I hate sounding like a broken record 😄 - I'm sorry about that!).

I've updated this (with a trival change). I think this is now good to go.

Can you rebase this on master -- it currently won't merge cleanly. That will give us another go around with Travis, too.

The only conflict is a trivial one in doc/users/whats_new.rst.

Tests look OK, with 1 error and 2 failures related to fonts; I don't think there is any relation to this PR.

I've squashed and rebased. Just waiting for travis-ci to do its thing & then I think it's good to go.

My only hesitation is that the PDF backend isn't matching Agg now. Have you had a chance to look into that, or should I?

I'm not sure what you're referring to. What is the issue/mismatch with PDF?

Thealpha_background test is turned off for PDF and there is a comment "PDF is wrong". I was just thinking maybe we should make it correct if possible before merging this.

As far as I can tell, the PDF being produced is correct. The ghostscript conversion of the PDF, however, insists on blending the PDF data with a white page background colour (eliminating all transparency in the process), and I've been unable to figure out a way around that. (The options that do support transparency seem to consider it an all-or-nothing matter.) I've also been unable to find a way specify page colour within the PDF, and I suspect there isn't one.

Converting the PDF to PNG using Inkscape gives a PNG with the correct colours and alpha values.

Ah, I see. Well, maybe we should just update the comment then, to
something like: "Ghostscript does not preserve the background color, so it
is not useful to test this with PDF".

On Apr 29, 2013 9:06 PM, "Wes Campaigne"notifications@github.com wrote:

As far as I can tell, the PDF being produced is correct. The ghostscript
conversion of the PDF, however, insists on blending the PDF data with a
white page background colour (eliminating all transparency in the process),
and I've been unable to figure out a way around that. (The options that do
support transparency seem to consider it an all-or-nothing matter.) I've
also been unable to find a way specify page colour within the PDF, and I
suspect there isn't one.

Converting the PDF to PNG using Inkscape gives a PNG with the correct
colours and alpha values.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/1868#issuecomment-17204384
.

I've just added that comment@mdboom in an appended commit. I always find it odd that I can modify the history of another author with git...

Woops, pushed the wrong branch. Fixed now.

Thanks for merging@efiring. Great work@Westacular - this is really nitty-gritty matplotlib. Thanks for your contribution!