Somebody on another forum put it better than I could: "this is a problem only on some images on the same roll as good scans." Same roll, same light, same settings — and half the frames need a nudge the other half do not.
That is not carelessness. It is structural, and it has a cause.
#What the software is actually solving for
A colour negative is three dye layers on an orange base. To read a dye you need to know how much light got through it relative to how much went in — a density needs a denominator. The denominator is the film base: the clear rebate around the frame.
If the base is guessed, estimated from the image, or picked off a border by eye, then every frame carries whatever error that guess made. And the guess is not the same on every frame, because the frames are not the same.
#Why per-frame correction hides it rather than fixing it
When the denominator is wrong, the picture still comes out — just shifted. So you nudge it back. Frame by frame, that works. Across a roll it does not, because you are correcting a measurement error with a taste adjustment, and the two do not track each other.
The frames that "worked first time" were the ones where the guess happened to be close.
#What measuring once per roll changes
The base is measured off the clear rebate, in density, once for the roll. Every frame on that roll is then read against the same denominator.
This does not make the colour better on any single frame. It makes the roll agree with itself — which is a different property, and the one that was actually missing.
Note
It also means a roll needs its rebate visible. If you crop to the image area in camera, there is nothing to measure and the app will say so rather than guess.
#What it costs to skip
Two measurements, both on a full 36-frame roll:
- Lighting with white ×1 instead of three narrow bands costs about five counts per channel.
- Running camera-blind — no measured profile for the camera — costs
[−5, −2.5, −2]counts, and the sign was uniform on 36 out of 36 frames.
A uniform sign is the interesting part. It is not noise you can average away; it is a bias that sits in the same direction on every frame.
I would rather publish those numbers than not. Nobody in this corner of the industry does, and a number you can check is worth more than an adjective.