[nuke] AgXc DRT investigation

[MrLixm]

This will act as a logbook for the further investigation of the AgXc DRT for Nuke that was release in v0.10.0.

I will note anything that I found that could help improve it.

[MrLixm]

print contrast

Test around the "print contrast" options. The initial release put it after the second tonescale but before it is linearized.

testA

put the Print contrast sigmoid parabolic after the linearisation.

• result: the pivot need to be adjusted to 0.355 to get a similar match with the original 0.65 pivot. A small difference is noticeable in the shoulder.
• conclusion: I find it more intuitive as putting 0.18 as pivot is clear that anything midgrey will be neutral

test B

sandwich the sigmoid parabolic between Log Shapers (still after linearization)

• result: create some weird bumps in the 2D plots around the toe part.
• conclusion: forget about it

[MrLixm]

outset

The outset is currently behind a linearize sandwich and applied after the first tonescale. The amount is driven by a simple lerp (NodeWrapper node) between the input of the outset and its output.

testA

remove the linearized sandwich

• result: the outset at maximum amount produces much softer result and avoid skews on problematic areas
• conclusion: that should be the new default

testB

Test the effect of the second Inset (part for the second tonescale section).

For that we disable the second log shaper + tonescale.

The second inset has default value of inset:0.4, r: 0.06, g: 0, b: 0.06 while first inset has values of inset: 0.4, r: 0.125, g: 0, b: 0.15

• result:
  • when the outset is at maximum amount, removing the second inset reintroduces skews around problematic areas
  • when the outset is at maximum amount, reducing the second inset with a lower valuer than the first inset reintroduces skews around problematic areas
  • swapping the outset and the second inset of order when the later has the same values as the first inset doesn't produce any difference

^ without second inset

^ with second inset, same global inset values as first inset (r,g,b inset are different)

^ removed the outset and the second inset

^ changed the second inset so its value perfectly match the first inset. It's a perfect match as if we remove outset+second inset.

• conclusions: the second inset and the outset are actually not useful when the second inset match the first inset. An interesting result can be produced if the second inset use different r,g,b inset

[MrLixm]

luminance compensation

Something I wanted to tackle since the beginning. I'm not even sure those words are accurate to describe the issue but again keeping in lack of better. This is of course in reference to the Eary and other Blender-Artist folks implementation: https://github.com/EaryChow/AgX_LUT_Gen/blob/main/luminance_compenstation_bt2020.py

While experimenting with grading under the AgXcDRT I stumbled on a cool side effect of operating on chroma in isolation of luma. Let me showcase it :

^ default AgXcDRT with no compensation

^ default AgXcDRT with compensation applied

The effect is most visible on the green dragon cg render.

The algorithm consist in :

The algorithm is as follows :

1. find luma using R-G-B weights computed from colorspace primaries and whitepoint
2. extract chroma by dividing luma
3. apply +1 offset on chroma
4. apply 0.5 power on chroma
5. apply 0.25 gain on luma to compensate for chroma change
6. multiplying back chroma with luma to recombine image
7. lerping 6. with luma calculated in 1. to only apply changes on higher values

Again no idea why it works, but it does.

issues

The observed issue with the current implementation are that it fix high values, but not low values as you can see on the green dragon. This can be addressed by removing the lerp with luma, but which create highly natural "brilliant" colors, especially the blue.