Calvin Williamson wrote:
> Eg He shows that if you want to downsample and then composite an image over
> a background, thats different from compositing first and then downsampling
> if you are using un-premultiplied images (you get different results that is).
> But with pre-multiplied images you get the same answer with either order.
<technical correctness sippage, flattening of color components snippage...>
> So it depends on how you work with your images. You might not be
> using the "compositing engine" part of the paint program at all.
> You might need to edit/paint the RGBs and Alpha for a layer more
> or less independently. In this case Alpha=0 and Color!=0
> channels might be fine. Maybe the correct Alpha channel isnt even there
> yet. Or maybe you are going to paint the alpha channel yourself.
Layer trees saw some play at GimpCon (http://www.gimp.org/gimpcon)
where the unidirectional layer stack differentiates into something with
branches. Users could composite/paint on one branch without disturbing
others - I think the metaphor to make dependency trees was made. Anyhow,
branches can distinguish themselves from others by having different compositional
rules. Users with compositional aims may wish to work with a pre multiplied
layer branch, knowing that certain tools are just not appropriate for that milieu.
Conversely users with painterly aims, who think in terms of alpha masking,
would find an unmultiplied layer branch to their liking - and mask-oriented tools
like unerase become readily functional.
To my mind, there is no absolute "rightness" to premultiplied layers,
they are "right" only when they help users achieve particular aims. If Gimp (2.0)
aims at being a general paint and compositing program, then its core needs to
support both approaches to using alpha, and make the distinction of different kinds
of layers visible in the user interface (becoming another beast in a layer menagerie
that includes Gimp classic layers, text layers, structured vector graphic layers...)
Be good, be well