TGA image supporting applications

[Alien]

This test was made to understand which applications create and save transparent and semi transparent TGA files correctly so no fixtrans is required. I tested image with opaque borders, transparent background and semi transparent brushes. PNG version:


Paint.NET:
Transparency is saved correctly, header is always saved wrong.

GIMP:
Transparency is saved correctly, header is preserved from already existing file (do not fix wrong one) or if the new file is created by GIMP, correct header is appended.

ImageMagick:
Preserves header (do not fix wrong one).

XnView:
Fixes and writes correct header.

IrfanView:
Fixes and writes correct header.

ArtGem:
Always writes wrong header. Transparency is saved correctly.

Painter X:
Transparency and header are saved correctly.

Photoshop CS4:
Always writes wrong header. Didn't know any convenient way to get semi transparent images working.


Addtional notes:
Fixtrans seems to mess RGB color space. It is seen both in XnView and Painter X.
RGB color space in Painter X:

RGB color space after fixtrans'ed conversion:


Need to look at it deeper, though.

[divVerent]

Exactly that is fixtrans's purpose. Many applications simply store black (or another color you do not want, or a color that once was there before) in unallocated pixels, or do not allow changing the color of fully transparent pixels without having to make them nontransparent first.

However, the resampling algorithms used in the graphics cards linearily interpolate BOTH alpha AND RGB data. So a pixel with alpha value 0 still has influence when a neighbor pixel has alpha larger than 0. This tends to cause weird borders around images. This problem gets even worse when the engine creates mipmaps, because that uses the exact same algorithm.

The only thing fixtrans does is replacing the RGB color of alpha 0 pixels by some pixel nearby, keeping alpha at 0, so the artifacts from the wrong resampling (that all GPUs, and DP use, because it is much faster than doing it right) disappear.

So the "RGB color space" of your test image is actually wrong: the grey part of your figure has black alpha-zero pixels surrounding it, which causes black outline artifacts when rendering. Fixtrans is there to fix exactly that.

That fixtrans also converts from PNG to TGA, or writes good TGA headers, is merely a side effect.

[Alien]

Test image is wrong, cause I tested TGA not PNG file. This is merely a converted TGA into PNG (converter was bad then).
Here, are 2 files which I compared. Both are reported as good in fixtrans. Yet, RGB is shown as different. So maybe fixtrans do not detect smth then:
http://www.mediafire.com/?ftimgtsk85f
http://www.mediafire.com/?zitdm30dm42

Exactly that is fixtrans's purpose. Many applications simply store black (or another color you do not want, or a color that once was there before) in unallocated pixels, or do not allow changing the color of fully transparent pixels without having to make them nontransparent first.

However, the resampling algorithms used in the graphics cards linearily interpolate BOTH alpha AND RGB data. So a pixel with alpha value 0 still has influence when a neighbor pixel has alpha larger than 0. This tends to cause weird borders around images. This problem gets even worse when the engine creates mipmaps, because that uses the exact same algorithm.

The only thing fixtrans does is replacing the RGB color of alpha 0 pixels by some pixel nearby, keeping alpha at 0, so the artifacts from the wrong resampling (that all GPUs, and DP use, because it is much faster than doing it right) disappear.

So the "RGB color space" of your test image is actually wrong: the grey part of your figure has black alpha-zero pixels surrounding it, which causes black outline artifacts when rendering. Fixtrans is there to fix exactly that.

That fixtrans also converts from PNG to TGA, or writes good TGA headers, is merely a side effect.

Why does it matter which colour fully transparent pixel uses? If it is fully transparent, you can't see it anyway. Or is it something wrong in graphic cards that neighbour pixel color has an effect? How does artifacts disappear using wrong resampling?

[divVerent]

Sure, fixtrans is not even trying to do that, I told you what it DOES.

[Alien]

EDITED above post. And yet shouldn't fixtrans fix it too?

[divVerent]

Fix what? What is broken? It can't make stuff transparent that isn't. That was never its purpose.

And yes, GPUs DO use the wrong algorithm when interpolating, which fixtrans is meant to fix.

E.g.

Pixel (0,0) = (red, alpha=1)
Pixel (1,0) = (red, alpha=1)
Pixel (0,1) = (red, alpha=1)
Pixel (1,1) = (green, alpha=0)

If you try to render pixel (0.5,0.5) (which happens when using e.g. GL_LINEAR texture mode), you will get:

(75% red, 25% green, alpha=0.75)

That's what happens then... and shows how the green pixel gets an influence.

[Alien]

Fix what? What is broken? It can't make stuff transparent that isn't. That was never its purpose.

Make image ABOVE look like correct one (BELOW) - fix black pixels (black in photoshop, paint x means transparency).


Thanks for info about cards.

In the wrong one pic I posted, black pixels would get influence.

[divVerent]

It does do that just fine. Unless you are referring to another image "above".

[Alien]

I'm referring to this bad image:

[divVerent]

fixtrans can't turn a non-transparent image into a transparent one, if you mean that. DP still loads the image with its normal TGA loading routine.