Light quality, not just light direction

Diffuse light gives steadier scores. It produces soft, gradual shadows that barely move when you shift the camera or the subject by a little, so retakes land closer together. Hard light produces sharp shadow edges that move with every small change, and a tool reads each one as contrast it has to account for.

Direction is one thing, quality is another

Light direction is where the light comes from. Quality is how it is delivered: from a small, bright source that casts crisp shadows, or from a large, soft one that wraps around the subject.

You can have the same direction with either quality. A bare bulb to the left and a large window to the left are the same direction and give very different images. So the direction post does not settle this, and neither does the comparison of flash and window light, which is about one particular hard-versus-soft pair, where distance and timing matter more.

What each does to an image

Under hard light, shadows have well-defined edges, and those edges sit in specific places. Nudge the subject, the lamp or the phone by a centimetre and the edges shift. Highlights are small and bright, and the transition from lit to shaded happens over a few pixels.

Under diffuse light, shadows are gradual and low-contrast, and small movements change them little. Detail is distributed more evenly, and highlights are broader and less likely to clip.

Google Research's work on adding light to portraits makes the underlying idea explicit: the direction of the light relative to the face's geometry determines how bright each part appears (Google Research, 2020). Hard light makes that dependence sharp, so every small geometry change shows. Diffuse light spreads it out, so there is less to react to.

What the tool reads

An image model scores what it can see in the pixels. A hard shadow is a high-contrast feature, and depending on the rubric it may be read as shape, depth, texture or damage. The same shadow in a different position reads differently.

The tool has no way to know the shadow came from a bulb rather than from the subject. So the extra contrast becomes part of the number, and it varies with every small variation in the setup. That variation is the spread you see between retakes.

A standard setup to compare them

You do not need equipment beyond what you have.

1. Fix the phone on a tripod or a stable support, with the lens, distance and settings you use for every submission.
2. Pick one light direction and keep it.
3. Take five photos with a bare, small, bright source, such as a single lamp without a shade or a phone torch.
4. Take five photos with the same source softened: a lampshade, a sheet of white paper or fabric in front, or a window on an overcast day.
5. Submit all ten, in a mixed order, to one tool.
6. Compare the spread in each group, meaning the gap between the highest and lowest of five.

The question is whether the spread narrows under the soft source. Do not worry that the averages differ. The averages answering "which scores higher" is the wrong question; spread and consistency are what you are measuring.

What to expect, and what not to

Expect a tighter group under diffuse light, though the size of the effect depends on the tool, and some tools are less sensitive to shadow than others. Expect the hard-light group to contain an occasional outlier when a shadow edge fell somewhere unlucky. Do not expect any particular score difference between groups, and do not trust a single pair. One retake is wild covers what to do with an outlier, and five retakes, one number covers why five is a sensible group size.

If the spread is the same under both, your tool is insensitive to light quality, and you can relax on this variable. That is a result.

Practical choices

Diffuse light is easy to get: a large window on an overcast day, a lamp bounced off a white wall or ceiling, or a lamp behind a sheet. It is also easy to repeat, which is the property that matters. Hard light is not wrong, but it demands that the source, the distance and the subject position be reproduced very precisely, and most home setups cannot do that.

Neighbours

How the model inside a tool responds to contrast and shadow is a design question, and aipenis.com's accuracy hub is the place for it. Tape-based measuring methods also care about lighting for readability, and measuremycock.com's method hub covers that discipline. How human reviewers handle harsh or flattering lighting in what they are shown belongs to ratepenis.com's judges hub. Rate Cock lists a breakdown with each result, which makes it easier to see whether a lighting change moved one axis or all of them.