I was wrong about my own files

I put out a guide saying a DCTL is more accurate than a LUT. It's the sort of thing everyone in colour repeats, and I repeated it, and I had a measurement I thought backed it up.

Then I did something I should have done first. I put my own RED LUT side by side with the exact transform and looked at them.

I couldn't see it

Not on ordinary shots. Not on the stress cases I picked specifically to break it. I went looking for the difference I'd just published a claim about and I couldn't find it with my eyes.

Which proves nothing. My eyes aren't evidence, and the whole reason this field has numbers is that everyone's eyes lie to them. So I stopped looking and asked for a number instead.

The number disagreed with my own page

Two hundred thousand samples across the colour cube. Then again, on three real frames, measured both ways to be sure the method wasn't doing the work.

The result contradicted what I'd published. Not by a little.

So the guide got corrected. It now has a section headed Where the error actually lives, and the first thing it says is that it isn't where we first wrote. Leaving that heading in was deliberate. A quiet edit would have been easier and worth less.

What it actually says

On the files I ship, no pixel lands above the threshold where a person can tell the difference. The worst one measures 0.84 against a threshold of 1.0.

On the smaller grid that most packs in this category ship, about 6 percent of pixels cross it. Worst case 1.36.

That's a real difference and a modest one. It's nothing like the difference I'd implied when I wrote that a DCTL beats a LUT.

The bit I didn't expect

The error is larger near neutral than it is in saturated colour. 0.095 against 0.048.

That's backwards from how I'd have guessed, and it matters more than the size of it suggests, because near neutral is exactly where the eye is most sensitive. Skin, concrete, overcast sky, the grey card. All the places you'd notice.

The part I want to say plainly

I chose 65 point long before any of this. Before the DCTL existed, before there was an argument to win, and for no cleverer reason than that bigger seemed better and the file size didn't matter.

It would be a much better story if I'd reasoned my way there. I didn't. I got lucky, and then I got curious enough to check, and checking is the only part of that I'd defend.

If you want the full numbers and the method, they're in the guide.