The colour temperature variable

A phone camera does not know what colour the light in your room actually is. It guesses, on every single shot, and the guess is called white balance. Get the guess wrong and the whole image shifts warm or cool, which changes how tone and texture read to anything scoring the picture.

What auto white balance actually does

Most cameras run automatic white balance by default, which means the camera looks at the scene and picks a colour temperature it thinks is neutral. Warm indoor bulbs push it one way, overcast daylight pushes it the other, and mixed lighting - a lamp and a window in the same shot - gives the algorithm two competing answers and no clean one. The result is that the same room, photographed twice a few minutes apart, can come back with a visibly different cast. Nothing about the subject changed. The camera's guess did.

Why this matters for a rated submission

A rating tool is reading pixels, and colour cast changes what those pixels say before anything about proportion or symmetry gets involved. The geometry of what a model reads from an image is one input; colour is another, sitting underneath it and shifting the same texture cues a warmer or cooler direction depending on the cast. A cool cast can read as flatter and less detailed; a warm cast can read as more saturated than the surface actually is. The effect is measurable in general-purpose models too: Afifi and Brown (2019) showed that strong colour casts from incorrectly applied white balance significantly degrade deep network performance on image classification and segmentation. Neither is the tool being wrong. It is scoring the photograph it was given, and the photograph is not a neutral copy of the subject - a distinction the same as why a tape measurement and a photograph are not interchangeable.

What to do about it

Locking white balance is not necessary. Being consistent about it is. Shoot in the same light source each time, at the same rough time of day if you are relying on a window, and let auto white balance make roughly the same guess it made last time. That is a different fix from correcting a filter or an edit after the fact, which changes the image rather than the capture conditions and breaks comparability in a more obvious way.

Colour cast belongs on the same line as distance, angle and crop: a condition worth writing down next to the result, not a setting worth fighting with. A tool that publishes per-axis results, the way Rate Cock does on public entries, at least gives you somewhere to look if one axis moves after a lighting change and the others do not. A note that records the light source alongside the file is enough to catch a shift before it gets mistaken for a real change, and it costs nothing beyond the habit of writing it down. A human reader would notice a colour shift and discount it without thinking; a person reviewing a submission directly does this instinctively in a way a fixed rubric does not.