This is not about Grading

Musicvid wrote on 2/10/2019, 7:20 PM

It is about multilayer Compositing, as differentiated from simple grading.

It is a continuation of this thread, which was quite predictably derailed by a drive-by saboteur.

Basis and Structure

To test the narrow conclusions proposed in the following statement, which originated at SCS long ago, I limited my inquiry to compositing (not grading per se), and followed the conditions exactly as stated without attempting to qualify, influence, or embellish the outcomes. For that reason, I have banned myself from any further comment or participation in the discussion. Contribute what you will, even conjecture..As always, well-framed peer tests are welcome.

When using 8-bit input/output, the 32-bit floating point (video levels) setting can prevent banding from compositing that contains fades, feathered edges, or gradients.

Testing Setup

  • Background Image -- 8 bit RGB 444 1080p
  • Composite layer (generated) -- 16 bit RGBA 444 Gradient 480p
  • Compositing pixel formats -- 8 bit, Gamma 2.2, and 32 bit float (16 bit realized), Video Levels, Gamma 2,2.
  • Output -- 8 bit RGB 444 Uncompressed


  • The 640x480 16 bit RGBA generated composite was place above the 1920x1080 8 bit background in an 8 bit project and output, at Source Alpha. Gradient track opacity was reduced to 20%, further exposing an area of 0 density (full transparency) in the upper right corner. This was done to meet the criteria, "fades, feathered edges, or gradients."
  • Output was 8 bit RGB 444, meeting the criteria, "When using 8-bit input/output, ..." The 16 bit generated composite was chosen due to indigent native banding in the 8 bit generated version.
  • The two captured frames were superimposed on the timeline, and the Difference compositing mode was chosen, eliminating the need for hierarchical placement of the two tracks.

  • A separate enhanced-contrast frame capture was made at 12,800% dynamic range factor (128:1) by setting the Input range to [0, 0.004] (0-2 RGB) This effectively isolates vestigial noise in the bottom 0.4%% of the output. I'll throw up the PSNR later, which is designed to measure just that

Musicvid wrote on 6/7/2018, 8:01 PM

My big deal about the scientific method is excluding the Pygmalion Effect completely (5 up), not in designing tests that appear to support my conjectures.

That all said, your test ignores the collateral impact of Vegas' dirty pattern dither, because you did not follow through and render an 8 bit lossless file from your float project. I did that too, and found it completely negated any theoretical effects of doing what you hope to accomplish. That's the conservative version.

Here's the real world version:

My only lingering question is the one I first raised in relation to my first round of tests on banding vs. dithering collateral, ca. 2012 (?):

"Is the minimal reduction in banding (diagonal lines) in the composite gradient worth the effects of bit-level pattern dither noise resulting from downsampling in Vegas? Is the difference worth the added noise, time, preview slowdown, and risk of introducing incongruities into the mix, as were reported? "

I won't attempt to answer that out of healthy concern of retaliation on a public forum, but anyone choosing to try needs to remain conscious that the image directly above represents a 12,800% expansion of the bottom 0.4% of values.

That's a good place for me to leave off, and I don't plan on responding to PM questions either, except with a handful of established contacts (you know who your are). Have a ball, as I intended, and Sayonara!




Musicvid wrote on 2/14/2019, 9:19 AM


If we isolate the project dither noise in Vegas' timeline, we find it occupies the bottom 0-8 RGB, or 3.2% of the RGB range. If I switch to and from 32-bit float video levels 2.2, the appearance of the dither noise in the scopes and preview is no different. No difference either if I employ an 8- or 16-bit gradient as my compositing layer.

That's a lot of detail to cover up if you're already shooting at murky levels, and composite grading, whether or not one renders in 8- or 10-bit format.

Lacking total debunking of my test methods (and I will not rule it out), I would like to elevate this to a feature request to employ a different dither method that introduces less noise in the bottom 3.2% of post-graded levels.

After I've slept on this one more time, I intend to send Nick a zipped VP14 project with assets, for independent verification of my data.