[Bf-committers] Grading (WAS: Vignette node)

[Troy Sobotka]

On Mar 7, 2015 5:29 AM, "Xavier Thomas" <xavier.thomas.1980 at gmail.com>
wrote:

> What do you mean by no scopes in the compositor?

> I don't consider the scopes of the sequencer nice, I consider them
horrible
> and horribly slow; they really should be ported to openGL and reuse the
> code of the image scopes.

The scopes via UV are great, and I too wish the the Viewer were somehow
more central to viewing. Perhaps that fits in with Mr. Wilkin's and
Mr. Riakiotakis' work?

We have a relatively minor issue in that the scopes assume a hard coded
color space. I have a proposal to integrate an XYZ role to alleviate this,
but I have to get it written up and submitted to Sergey. Hopefully then we
can update the scopes to work against arbitrary reference spaces.

> I agree that the compositor is not the best place to do grading and I
think
> the sequencer would be better.

I would suggest that the VSE is sub-optimal on a number of fronts if
considered alone, with some significant potential upside. Of note:

A) The VSE was designed for realtime playback. As such, most of the VSE's
work has legacy low quality 8 bpc code paths.
B) Contemporary grading suites all leverage nodal compositing due to the
complexity of grades.

So in terms of a grading design, I might suggest:
A) Rework the VSE to use a fully “offline” system. That is, leverage the
shot based nature to streamline proxy work for the editing, blocking,
timing, and preview portion.
B) Rework the VSE to provide a conforming / grading / finishing “online”
mode that works with a compositor view. Likely another compositor mode that
allows the full 32 bit algorithms to work for the highest quality output.

Some elements that would need further fleshing out, but it would at least
appear to be the foundation for a decent system without significantly
altering the overarching existing design for Blender. An additional
compositor mode and some VSE glue.

The only stumbling block would be the need for the VSE's curves to accept
OCIO transforms. For example, if an artist draws a linear diagonal curve
for a transform, if the artist is working in a display referred space, the
diagonal line can be assumed to be a nonlinear representation. To get a
proper dissolve / fade / over, it would need to be inverted via the OCIO
transform and the composite would be performed on linearized data at
whatever the reference primaries are.

With respect,
TJS