Alpha blending issue

[frdbr]

Alpha blending doesn't work properly leading to ill-blended artifacts in transitions. See example in Kdenlive bug report. Alpha blending can't be done properly in 8 bits uint because gamma needs to be decoded for alpha to mean occlusion as in the Porter & Duff paper. More information on wikipedia and youtube.

A possible solution would be to:

1. Step one would be to use at least float16
2. Then undo gamma for every RGB
3. Then compose, then redo gamma on RGB

This report is compiled from information provided by @aurelienpierreeng here and here

[ddennedy]

This is a road map set of items, works correctly, in Movit, and is not limited to a few places. Please be aware that converting all processing to work in linear RGB will change the behavior of many color-oriented filters essentially breaking compatibility if not done correctly and thus existing projects should not benefit from a change here. Fixing it also has a big impact on performance and so must be coupled with something that works well with GPU. Then to keep performance good while still having a decently large toolset of effects means many things need to be ported. But there is vast landscape of GPU technologies, and there should always be a multi-threaded CPU fallback that is not GPU emulation. So, this is a whole big can of worms.

[ddennedy]

Fixed by #1159 for the next version.
It requires the app to opt into by supplying a new consumer property mlt_color_trc=linear. This adds a considerable performance overhead, which we hope to improve some, but I would not expect a lot. An app/user could do the interactive work in like a draft mode without this and then turn it on only for export render. Next, upon review of the result, if some things need more precise adjustment turn on linear within the app to fine tune and re-render. Similar to a workflow with preview scaling.