[BUG] Antialias subsampling handles subsample transparency wrong

Is it a bug? I've noticed it too but I think it's just interpolating the colors at the same time as the alpha channel. It would be nice to not have this though too...

Quote
Betis wrote: Is it a bug? I've noticed it too but I think it's just interpolating the colors at the same time as the alpha channel. It would be nice to not have this though too...

Bevel width is 0 (see settings on attached image), so there is no interpolation going on.

The red fringe issue is only there when antialiasing is on, so I believe the problem is that the output of the sub sampling is a simple average of all the sub samples collected. This means that sub samples with alpha = 0 will pollute the output like seen here.

The solution is very simple: just multiply the sub samples with their alpha before adding them up and divide the sum result by the sum of alpha and not sample count.

Quote
uberzev wrote: Does this issue remain after saving/exporting the image or only occur in the preview?

Yes, unfortunatly it does - see the explanation above.

Sorry, understanding your posts more clearly now, I can see how this could be handled correctly. +1

Cool - yeah it is not really a question about interpolation - and actually in this case not alpha blending either (I wish I could rename the topic and rewrite the intro: in this case it is all about how the subsampling works).

Basically it sums up to this: the alpha channel is a property of a color as well as red, green and blue channels. The blue channel logically enough controls the amount of blue light, red and green likewise. But the alpha is different: it controls all of them - zero alpha means 100% transparent, i.e. NO red, green or blue light *at all*.

That fact does not change just because you're doing subsampling or blurring or blending (and in some case even interpolation).

Quote
Skybase wrote: I thought this was just behavior and that if you need to get rid of those ugly lines, they can be premultiplied when imported to a different program. This basically happens with some of the 3D programs I use as well so I usually handle with pre-multiply since that's just been my workflow.

No, you can't fix that simply by importing the image as if the alpha channel were premultiplied (it will just make the edges more jaggy in the alpha regions). The sum is already polluted with the transparent colors, and there is no way to get that out again.

Quote
Sphinx. wrote: I wish I could rename the topic and rewrite the intro

What would you like to rename the topic to? I can modify the first post as you wish.
UPD: Modified per Sphinx's request.