2003.02.04 22:37 "Re: Pre-multiplied alpha", by Chris Cox
The tags in that image look right for a premultiplied image, but the pixels all have an alpha value of zero, even in the red pixels. Images with a color value > alpha value are sort of ill-defined and something tells me that may not be what you intended.
I don't have Photoshop handy (I'm at a linux box), Pixar's internal tools can show the image just fine, as can shake but the gimp sees the zero alpha and just seems to ignore the color. I do know that a few Photoshop revisions ago, it was common practice around here to run tiff files through a filter that clamped all the color channels to be <= alpha before sending them to Photoshop.
Because Photoshop internally does not use premultiplied alpha (for many reasons). We have to dematte the data as we bring it in. The dematting process clamps the "superluminous" colors.
Although I'm not sure why you had to clamp it yourself.
I'm not sure what Adobe assumes should be done with over-alpha colors, but in general such images but in practice, software handling of this kind of image is pretty varied depending on the assumptions of the authors.
If you have a package that understands high-dynamic range images it'll likely handle them fine, but a package which expects a normalized color range, (like I would assume Photoshop is), having color > alpha can break a whole lot of software assumptions.